Development of an imaging radar capable of detecting hidden obstacles for terrain
mapping on an autonomous off-road vehicle
Daniel Gusland
Thesis submitted for the degree of Master in Cybernetics
60 credits
Department of Technology Systems Faculty of mathematics and natural sciences
UNIVERSITY OF OSLO
Development of an imaging radar capable of detecting hidden obstacles for terrain
mapping on an autonomous off-road vehicle
Daniel Gusland
Development of an imaging radar capable of detecting hidden obstacles for terrain mapping on an autonomous off-road vehicle
http://www.duo.uio.no/
Printed: Norwegian Defence Research Establishment
Preface
Ten months ago I embarked on what has been my biggest academic journey yet. To satisfy the enthusiasm for the way computers perceive the world, I started the challenge of detecting hidden obstacles. Fueled by the excitement of the problem at hand, having never worked with radar beforehand. This thesis is the results of countless long days in the lab, thousands of pages read and many setbacks. To me it stands as a testimony that hard work and a positive attitude pays off in the end.
The work on this thesis has also served as an academic epiphany, being continually surprised of the cleverness of scientists and students worldwide. It is truly astonishing amounts of research available, even in the smallest of fields. I would like to dedicate thanks to Gregory L. Char- vat for writing “Small and Short Range Radar Systems”, which has been a truly great resource throughout the project.
I would to use this opportunity to acknowledge my advisor Dr. Børge Torvik. Thank you for all your help, support and enthusiasm throughout the project. Special thanks to Idar Dyrdal for your keen academic guidance, and for believing in me from i started writing my Bachelors thesis up until now. I would also like to acknowledge Lorns Bakstad for having faith in me, even when the results were absent. I acknowledge the Norwegian Defence Research Estab- lishent (FFI) for giving me the opportunity to work on extraordinarily interesting projects. I have received a lot of help and guidance at FFI, so I cannot write down everyone explicitly, but I am deeply grateful for all your help.
I would like to acknowledge my parents, for letting me pursue my interests and teaching me that hard work pays off. A very special thanks to my wife for putting up with many late nights and living with an at times obsessed husband. Thank you for all your support and love through- out the process. Lastly, but not least I would like to you, the reader, for your attention and I hope that you find the thesis interesting.
Summary
This thesis covers the development of an imaging radar capable of detecting hidden obstacles for terrain mapping for use on an autonomous off-road vehicle. Off-road environments pose several challenges for autonomous vehicles. This thesis focuses on one such problem; obstacles hidden in vegetation. The thesis starts with an extensive literature review of sensor techno- logies, such as daylight cameras, IR and spectral cameras and LiDAR sensors. All of these sensors have their strengths and weaknesses, but they all lack the ability to penetrate and detect truly hidden obstacles. This is where radar sensors are exceptional.
Radar sensors used for similar purposes are evaluated and analyzed. Large variations in con- figurations were found in the literature with regard to frequency and imaging techniques. Con- ventional mapping radars typically operate at high frequencies, which has proved to drastically reduce their ability to penetrate materials. To understand the large variations in configuration in both frequency and imaging techniques, literature reviews of each of these subjects were also undertaken.
To address the questions remaining from the literature review, a comprehensive radar systems analysis is presented. Starting with measurements and analysis of the attenuation of radar sig- nals in the relevant frequencies, before focusing the efforts on contrast between obstacles and vegetation. Concluding that a low frequency system operation in the 1-6 GHz interval might be able to detect hidden obstacles. For investigation of the latter a synthetic aperture radar measurement setup was developed, using a linear rail and a vector network analyzer. A short presentation of image formation algorithms is presented, with a particular focus on time-domain backprojection. Methods for increasing cross-range resolution with limited increase in system complexity were investigated and simulated. Concluding that a time-domain multiple-input multiple-output configuration is suitable for the application.
A demonstrator system is developed based on the previous analysis and simulations. All the ne- cessary considerations for developing a functional system based on simulations are presented, including system design and control. The demonstrator itself is built using mostly connector- ized commercially available components, but some system components had to be custom made for the application. Considerable performance limiting factors were addressed and mitigated.
Finally, the complete demonstrator system is tested in both controlled and complex environ- ments, starting with initial imaging results to ensure system functionality and compare meas- ured performance to the simulated results. The systems ability to detect obstacles in highly controlled environments is tested and found to be very good. The system was mounted on a test vehicle and combining the radar with the vehicles navigation enabled system testing in com- plex environments. Two test cases are presented; lightly occluded and fully occluded obstacles.
In the case with lightly occluded obstacles, the system performed well, detecting the obstacles unambiguously. The case with highly occluded obstacles posed a larger challenge. It is argued that the reason for the ambiguous obstacle detection in this case is a sub-optimal placement of the radar on the vehicle and the rough mapping technique implemented. It is believed that the radar is capable of detecting these obstacles if these issues are resolved.
Contents
Preface i
Summary iii
Table of Contents vii
List of Tables viii
List of Figures xii
Abbreviations xiii
1 Introduction 1
1.1 Research Objectives . . . 2
1.2 Thesis organization . . . 3
2 Research context 4 2.1 Alternative sensors . . . 5
2.1.1 Daylight cameras . . . 5
2.1.2 IR and spectral cameras . . . 5
2.1.3 LiDAR . . . 6
2.2 Radar systems . . . 7
2.3 Vegetation Attenuation . . . 8
2.4 Radar imaging systems and algorithms . . . 9
2.5 Summary . . . 10
3 Radar introduction 11 3.1 The concept of radar . . . 12
3.2 Radar Waveforms . . . 12
3.2.1 CW Waveform . . . 13
3.2.2 FMCW waveform . . . 14
3.3 Electromagnetic wave propagation . . . 16
3.3.1 Constitutive parameters . . . 17
3.3.2 Polarization . . . 17
3.3.3 Reflection and transmission . . . 18
3.4 Radar scattering . . . 18
3.5 Foliage Phenomenon . . . 19
3.6 Radar equipment used in this thesis . . . 20
3.6.1 Vector Network Analyzer . . . 20
3.6.2 Synthetic aperture radar rail . . . 20
3.6.3 Antennas . . . 22
4 Radar system analysis 23 4.1 Frequency analysis . . . 24
4.1.1 Attenuation . . . 24
4.1.2 Backscatter . . . 29
4.1.3 Contrast . . . 30
4.1.4 Summary . . . 36
4.2 Synthetic Aperture Radar . . . 37
4.2.1 Geometry . . . 38
4.2.2 Image formation algorithms . . . 39
4.2.3 Time domain backprojection . . . 40
4.3 System Simulation . . . 41
4.4 Resolution capabilities . . . 43
4.4.1 Conventional array resolution . . . 44
4.4.2 Cross-range resolution improvement . . . 46
4.5 Antenna configuration . . . 47
4.6 Summary . . . 54
5 Imaging radar demonstrator system 55 5.1 System description . . . 56
5.2 Transmitter . . . 56
5.2.1 Baseband signal generation . . . 56
5.2.2 Generating RF signal . . . 57
5.3 Switching Scheme . . . 58
5.4 Receiver . . . 59
5.5 IF-filter . . . 60
5.6 System control and digitization . . . 61
5.7 Waveform parameters . . . 62
5.8 Antenna coupling . . . 63
5.9 System testing . . . 64
5.9.1 Noise from switching power supply . . . 66
5.9.2 LO-delay reflections . . . 67
5.9.3 Dynamic Range evaluation . . . 68
5.10 Demonstrator construction . . . 70
5.11 Summary . . . 72
6 System testing and results 73 6.1 Initial imaging results . . . 74
6.1.1 Resolution measurements . . . 76
6.1.2 Calibration . . . 77
6.2 Obstacle contrast . . . 78
6.3 Navigation and mapping . . . 80
6.3.1 Navigation . . . 80
6.3.2 Mapping . . . 81
6.4 Complex scenes and occluded obstacles . . . 82
6.4.1 Test case 1 . . . 83
6.4.2 Test case 2 . . . 87
6.4.3 Discussion on the demonstrator system performance . . . 90
6.5 Summary . . . 90
7 Conclusion 91
Bibliography 92
Appendix
2.1 Summarization of the attenuation in corn-fields with experiments conducted by Ulaby et al. [1] . . . 9 3.1 Components of FMCW IF signal . . . 15 4.1 Radar frequency bands as presented in table 1.1 in “Principles of Modern Radar”
[2] . . . 25
List of Figures
1.1 LiDAR terrain analysis with vegetation present. . . 2 1.2 Vegetated dirt road, photo courtesy pxhere.com . . . 3 3.1 Time domain representation of a LFM chirp . . . 14 3.2 Illustration of the FMCW principle, showing the response of a single ideal point
scatterer. . . 16 3.3 Waves radiated from a source and approximated by a plane wave in the far-field. 16 3.4 Illustration of the polarization of the electric and magnetic field. Adapted from
Ulaby [3] . . . 17 3.5 Illustration of the scattering of an incident beam by a target, some vector graph-
ics provided by vecteezy.com. . . 18 3.6 Illustration of a multi-path reflection from a rough surface. . . 18 3.7 Description of structural classes of vegetation as described by Dobson [4], some
vector graphics provided by vecteezy.com. . . 19 3.8 High level diagram showing the components of the SAR rail. The host computer
is connected via Ethernet to the VNA and USB to the Arduino micro controller.
Using a python script on the host computer, the data is collected and the antenna positions recorded. . . 21 3.9 Image showing the SAR rail setup . . . 21 3.10 Flowchart of the SAR rail operation . . . 21 3.11 Backprojected image of the FFI logo spelled in nails, the measurements were
collected using the SAR rail utilizing 13 GHz bandwidth and horizontal polar- ization. Notice the footsteps surrounding the scene, caused by small variations in the grass when placing the nails. . . 22 4.1 Image of the one-way transmission measurement setup and result. . . 26 4.2 Conceptual drawing of the measurement setup to a relative scale. . . 26 4.3 Images showing the measurement setup of the two-way attenuation measurements. 27 4.4 Results from attenuation using horizontal and vertical polarization for the meas-
urements setup described in Section 4.1.1 . . . 28 4.5 Horizontal and vertical attenuation for the two-way measurement setup aver-
aged over all heights. . . 29 4.6 Range profiles with a reflector placed at 6-7 meters and vegetation present at
2-5 meters . . . 30 4.7 Pictures showing the collection of relevant obstacles and the measurement setup
for the RCS measurements. . . 31
view overlaid with a SAR image acquired during the measurements. . . 33
4.10 SAR image showing the target placement and measurement patches for the con- trast measurements. This image was processed using a center frequency of 3 GHz and a bandwidth of 1 GHz. . . 34
4.11 SAR images at several center frequencies before and after vegetation removal. . 35
4.12 Target contrast relative to clutter as a function of frequency . . . 36
4.13 Illustration of the forward-looking imaging geometry, some vector graphics provided by vecteezy.com. . . 37
4.14 Figures illustrating the change in range to target with respect to antenna position. 38 4.15 Illustration of the imaging SAR bistatic geometry . . . 39
4.16 Pseudo code for the implementation of time-domain backprojection in Matlab . 41 4.17 Simulated phase history with 10 m monostatic configuration, scatterers located at 2, 5, 10 and 20 m down-range. . . 42
4.18 Image created using time-domain backprojection on the phase history data shown in Figure 4.17. Notice the range dependent cross-range resolution. . . 43
4.19 Range resolution as a function of bandwidth . . . 44
4.20 Receive antenna array with 8 fully polarimetric channels. . . 44
4.21 Cross range resolution throughout a scene using a conventional array of 8 re- ceive elements. . . 45
4.22 Cross range resolution throughout a scene using a conventional array of 28 re- ceive elements. . . 46
4.23 Cross-range resolution as a function of range for a conventional array versus a monostatic SAR of the same length. . . 48
4.24 Full PSR and Cross-range PSR of a point scatterer at 10 m with different an- tenna configurations . . . 48
4.25 Antenna configurations for the initial resolution analysis. . . 49
4.26 Cross-range PSR of a point scatterer at 10 m with different antenna configurations 49 4.27 PSR of pointscatterer in Figure 4.28. . . 50
4.28 Imaged pointscatterer with bistatic and monostatic antenna configurations, show- ing a clear difference in sidelobes. . . 50
4.29 Illustration of the virtual phase centers for a bistatic mimo antenna. . . 51
4.30 PSR of point targets with different antenna configurations with and without tapering functions applied. . . 51
4.31 Illustration showing the receive array, the optimal transmit positions and their corresponding phase centers. . . 52
4.32 Full PSR and cross-range PSR of a point scatterer at 10 m with different antenna configurations . . . 52
4.33 Comparison of images created using a conventional array versus four transmit antennas. Scene contains 12 ideal point scatterers. . . 53
4.34 Illustration of the final antenna configuration . . . 53
5.1 Simple explanation of an FMCW radar system, some vector graphics provided by vecteezy.com. . . 56
5.2 Measurement of the DDS output before and after filtering . . . 57
5.3 Block diagram of the transmitter configuration . . . 58
5.4 Output of the RF mixer before and after filtering. . . 58
5.5 Switching scheme relative to the scene. . . 59
5.6 Switching scheme and input to the RF switch. . . 59
5.7 Block diagram of one receiver channel . . . 60
5.8 Circuit diagram of the IF filter simulation model. . . 60
5.9 Simulated plot of the IF-filter amplitude response with respect to frequency. The frequency response compensates for range. . . 61
5.10 Image of the final manufactured and assembled PCB. . . 61
5.11 Radar system overview and timing . . . 62
5.12 Flowchart of radar system setup and sampling . . . 62
5.13 Explanation of the DDS ramp generator [5] . . . 63
5.14 Illustration of antenna configurations to decrease antenna coupling. . . 63
5.15 Antenna setup with vertical spacing and RAM . . . 64
5.16 Measured antenna isolation using RAM of different lengths. . . 64
5.17 Free space measurements with a reflector at approximately 8 meters using a single sweep. . . 65
5.18 Free space measurements with a reflector at approximately 8 meters. Integration of 10.000 pulses. . . 66
5.19 Range-doppler plots before and after PSU noise has been removed. . . 67
5.20 Point response using a delay-line measurement, showing the improvements to the point response with different LO setups on the mixer. . . 68
5.21 Plot showing mixer intermodulation products. The peak at approximately 310 kHz, whereas the other peaks are a combination of harmonic responses and intermodulation products. . . 69
5.22 3D mechanical drawing of the demonstrator system . . . 70
5.23 Cardboard Prototype and the parts of the RF assembly. . . 70
5.24 Radar front end with signal cables, spigot mounting option, receiving antenna array array and two transmit antennas. . . 71
5.25 Inside view of the radar frontend. Most dominant is the filter-LNA-mixer as- sembly seen in the middle of the frontend. . . 71
5.26 Top layer of the TX rack unit showing . . . 71
5.27 Bottom layer of the signal generation rack unit, containing the RF chain and power supply units. . . 71
6.1 OLAV platform in heavy vegetation with radar mounted in front. . . 73
6.2 Measurement setup and scene for the first imaging experiment . . . 74
6.3 Flowchart showing the image creation pipeline. . . 75
6.4 Comparing backprojected images using respectively four transmit antennas versus one transmit antenna. . . 75
6.5 Radar image using four transmit antennas overlaid aerial photo of the area. . . . 76
6.6 Imaged reflector using 1TX antenna versus 4 TX antennas . . . 76
6.7 PSR of a trihedral reflector using 1 and 4 TX antennas, plotted alongside a 3 dB line . . . 77
6.8 Imaged scene of the contrast measurements of a barrel. . . 78
6.9 Target contrast to asphalt using ROLAV. . . 79
6.10 Bar chart showing target contrast relative to asphalt for a variety of obstacles. . 79
6.11 Explanation of the navigation of a local map using a flat earth approximation, some vector graphics provided by vecteezy.com. . . 80
6.12 Flowchart of the radar mapping procedure. . . 81
6.15 Snapshots of the obstacles from the vehicles view, taken from the daylight cam- eras mounted on the vehicle. From left to right, the obstacles are a small rock, large rock, concrete support, stub and a paint can respectively. . . 83 6.16 Radar map of the scene in test case 1. Images are created using backprojection
and mapped using the vehicles navigation and an iterative multi-look technique.
The image clearly shows the four obstacles, as well as other potential obstacles that have not been investigated . . . 84 6.17 Radar map from case one with obstacles present, overlaid an aerial photo. . . . 85 6.18 Radar map from case one without obstacles present, overlaid an aerial photo. . 85 6.19 A cropped part of the obstacle positions before and after obstacles were positioned. 86 6.20 Bar chart showing the difference in amplitude with and without an obstacle
present. . . 86 6.21 Snapshots of the obstacles from the vehicles view, taken from the daylight cam-
eras mounted on the vehicles. From left to right, the obstacles are a small rock, large rock, concrete support, stub and a paint can respectively. . . 87 6.22 Radar map from case two with obstacles present, overlaid an aerial photo. . . . 88 6.23 Radar map from case two without obstacles present, overlaid an aerial photo. . 88 6.24 A cropped part of the obstacle positions before and after obstacles were positioned. 89 6.25 Bar chart showing the difference in amplitude with and without an obstacle
present. . . 89
Abbreviations
COTS Commercial Off-The-Shelf
CW Continuous Wave
DDS Direct Digital Synthesis
EM Electromagnetic
FFI Norwegian Defence Research Establishent FMCW Frequency-Modulated Continuous Wave GPS Global Positioning System
IF Intermediate Frequency IR Infrared Radiation
LiDAR LIght Detection And Ranging
LO Local Oscillator
LFM Linear Frequency Modulation MIMO Multiple-Input and Multiple-Output RANSAC RAndom SAmpling Consensus RCS Radar Cross Section
RF Radio Frequency
RMA Range Migration Algorithm
ROLAV Radar on an Off-road Light Autonomous Vehicle SAR Synthetic Aperture Radar
SNR Signal to Noise Ratio TDM Time-Division Multiplexing
TX Transmit
UGV Unmanned Ground Vehicle VNA Vector Network Analyzer
Chapter 1
Introduction
Norwegian Defence Research Establishent (FFI) has conducted research on autonomous under- water vehicles and drones for years. Lately they have started researching autonomous surface vehicles. For this purpose a platform has been acquired to demonstrate the capabilities of an Unmanned Ground Vehicle (UGV). A considerable challenge in off-road conditions is that the surface can be covered by vegetation. The problem in it self is a dual problem; the vegetation might hide a true obstacle or the vegetation in itself might be non traversable.
It is interesting to investigate the candidate sensor technologies for detecting these obstacles.
Radar systems are particularly appealing, due to their proven ability to penetrate materials as demonstrated by through-the-wall imaging systems and ground penetrating radar. Having a radar system on an autonomous vehicle also introduces the potential to detect, locate, track and even classify incoming unmanned aerial vehicles.
There is evidently a need for a system that can identify obstacles hidden in vegetation to in- crease the maneuverability of the vehicle. The existing LIght Detection And Ranging (LiDAR) obstacle detection methods does not discriminate between vegetation and obstacles, as seen by the terrain analysis in Figure 1.1. The example shows that the light vegetation to the right in the image in Figure 1.1a, is interpreted as obstacles as shown in Figure 1.1b and 1.1c. Another example of an obviously drivable road is shown in Figure 1.2. Based on this LiDAR terrain assessment, this road would most likely be characterized as non-traversable.
The relevant obstacles can be separated into four categories: positive obstacles, like trees, poles, rocks and boulders; negative obstacles such as holes, ditches and non-load bearing surfaces;
moving obstacles like vehicles, personnel and animals and difficult terrain like slopes, water, mud etc. To complicate the classification, examples from each of these categories may also be traversable or non-traversable depending on the vehicles capabilities. The traversability of some obstacles can also be dependent on the situation. It is desirable that the system eventually should detect non-traversable positive obstacles like large rocks, tree stubs and poles. In this thesis, the aim is limited to positive obstacles.
(a)Image of the scene
(b)obstacle-colored pointcloud (c)Occupancy grid Figure 1.1:LiDAR terrain analysis with vegetation present.
1.1 Research Objectives
Designing a system capable of detecting obstacles occluded or otherwise hidden from view is a highly complex task. There is an obvious question as to the possibility of creating such a sys- tem. The research question fueling this thesis is therefore: “Is it possible to create a imaging radar prototype that is able to penetrate vegetation and foliage, whilst at the same time be able to detect relevant obstacles to augment the perception beyond laser and optical sensors?”
While a detailed review of the available literature and simulations might indicate the answer to this question, the best way to answer it is to develop such a system and demonstrate its capab- ilities. This thesis covers the development of such a demonstrator systems, with the following research objectives:
• Review alternative sensors
• Quantify the attenuation and contrast of relevant obstacles at different frequencies using available literature and measurements.
1.2 Thesis organization
Figure 1.2:Vegetated dirt road, photo courtesy pxhere.com
• Conclude with a radar configuration based on a literature review, simulations and fre- quency analysis.
• Build and test the simulated system.
• Demonstrate system performance in controlled and complex scenes.
1.2 Thesis organization
The thesis covers the development of the system chronologically, which conversely follows the research objectives. Chapter 2 gives a detailed review of the research context, starting with al- ternative sensors before continuing with radar systems and a detailed review of both attenuation and radar imaging. Chapter 3 lays the ground work for the general understanding of how a radar works. It gives a short review of the concept of radar, a mathematical review of the Frequency- Modulated Continuous Wave (FMCW) waveform which will be used extensively throughout the thesis. Continuing with the basics of electromagnetic wave propagation, attenuation, scat- tering and foliage phenomenon. Chapter 3 is rounded off with a quick explanation of the radar equipment used in the thesis.
Chapter 4 contains the frequency analysis focusing on both attenuation and target contrast.
Followed by a detailed review of Synthetic Aperture Radar (SAR) processing focused on time- domain backprojection, before continuing with radar system simulation to verify the imaging algorithm and test resolution calculations. The last part of the chapter is dedicated to antenna configurations, analyzed using simulations. Chapter 5 covers the actual construction of the demonstrator and the analysis needed to create a working system based on the theoretical model described in the previous chapter. It covers the basic radar system architecture and the basics of the system control and initial testing to verify performance. All these chapters accumulate in Chapter 6 where the final system is presented and tested in both controlled and complex en- vironments. The final discussions and conclusions are in Chapter 7, which also point to future work.
Chapter 2
Research context
There is a large amount of research on perception for autonomous vehicles. Most of the avail- able literature focuses on urban environments and understanding situations, detecting cars, people walking, cycling and other hazards. Some autonomous systems are already available in commercial vehicles; automated cruise control, automatic emergency braking etc. These fea- tures primarily use cameras and/or radars to achieve the desired functionality.
There are several challenges with driving in off-road environments compared to urban envir- onments. The vehicle has to decide if the surface is load bearing and smooth enough. In addition, it has to decide whether obstacles are harmful to the vehicle or not. There are several approaches to finding obstacles hidden in vegetation. Some of the deciding factors for choosing a technology are the system complexity, penetration depth, probability of detection and range.
The first part of this chapter contains a short review of alternative sensors. Continuing with a review of similar radar systems. At the end of the chapter, there is a review of general radar imaging and the effect of vegetation on microwaves.
2.1 Alternative sensors
2.1 Alternative sensors
There are several other sensors appropriate for detecting obstacles. Daylight cameras are an obvious and affordable solutions, as well as Infrared Radiation (IR) and to some degree spectral cameras. LiDAR sensors are less affordable, but are highly capable sensors for autonomous navigation.
2.1.1 Daylight cameras
The existing research on using cameras relevant to the research question mostly focus on de- tecting vegetation, as cameras cannot penetrate vegetation and therefore cannot determine what is truly hidden. There are several approaches. Huarguess and Larson addressed the issue of separating man-made objects from vegetation [6], as an effort of a passive way to detect veget- ation before further classification. They succeeded in detecting vegetation, but they were not able to detect man-made objects robustly. The most applicable work using cameras starts with Talukder et al. [7], where the authors redefine obstacle detection for off-road environments us- ing stereo cameras. Previous attempts had mostly used plane estimation methods like RAndom SAmpling Consensus (RANSAC), a method for estimating planes in datasets containing a large amount of outliers, but Talukder et al. defined an obstacle using simple rules for the neighboring environment of pixels, making the algorithms more robust to slopes and uneven terrain. The method developed by Talukder has been improved by Mark et al. [8], where they improve the clustering and by Dos Santos et al. [9] by adding a feature extractor and a bayesian classifier to classify obstacles such as people, bushes, animals and grass of different heights. Their work shows promising results in the examples shown in the paper. In agricultural environments, a change-detection method using stereo cameras has been proposed by Patrick Ross et al. [10], overcoming the shortcomings of a stereo camera in ambiguous environments by reconstructing only the parts of the image that have changed, before doing obstacle detection.
Camera solutions are affordable and accessible, but they lack the ability to penetrate relev- ant materials and therefore cannot estimate the presence of an actual hidden obstacle. However, cameras can be used to detect vegetation, visible or slightly occluded obstacles.
2.1.2 IR and spectral cameras
IR and spectral cameras capture specific parts of the electromagnetic spectrum and will in most cases extract different information than conventional daylight-cameras. There are several prom- ising results in detection and classification of vegetation using IR-cameras. Bradley et al. [11]
used a relationship between red and near-infrared reflectance to detect chlorophyll-rich veget- ation, which is often non-dense vegetation. Using this technique the authors were able to dis- criminate light vegetation and separate lightly occluded obstacles. Rankin and a team from Jet Propulsion Lab [12] were able to use thermal IR-cameras to discriminate vegetation, soil, water, mud and personnel. IR-cameras also have the ability to see through a certain amount of fog and smoke. Given these capabilities, IR cameras are a purposeful addition to a UGV sensor suite, but still lack the ability to penetrate vegetation.
2.1.3 LiDAR
LiDAR sensors measures the distance to an object by illuminating the target with a laser and measuring the reflected pulses. LiDAR sensors are popular for autonomous ground vehicle platforms as they provide a high-resolution 3D representation of the surrounding environment.
Most of the sensors are rotating laser scanners with one or more scanning lasers with a beam- width of a few milliradians. This does not give the sensor the ability to penetrate vegetation, however, it does enable the beam to go in between vegetation. LiDAR technology has been used for tree height, crown and biomass estimation from airborne platforms. Hopkins [13] studied the influence of beam divergence and flying altitude with regards to canopy penetration. The results of this study is confirmed by Wieser et al. [14], where they compare different airborne laser scanning configurations. It is found in this study that a low beam divergence allows for good penetration through canopies.
LiDAR has also been used on ground based platforms. Steinvall et al. at FOI (Swedish Defense Research Agency) studied the performance of LiDAR for imaging of concealed targets [15–17]
and testing different configurations. They were able to successfully characterize targets like vehicles and persons concealed by vegetation, glass and camouflage. The common factor of their testing is that the concealed obstacle is further away than a pulse-length from its con- cealment, simplifying the signal processing. Hollinger et al. [18, 19] investigate this further, concluding that scanning LiDARs are the best option. However, the false alarm rates are high, indicating the benefit of an additional sensor.
Despite its inherit limitations, LiDAR has been used to find partially or highly occluded obstacles with successful results. Matthies and Castano [20] address the problem, using the idea that LiDAR returns from an obstacle is continuous, whereas vegetation is discontinuous. Later Matthies et al. concluded that LiDAR-based obstacle detection can detect obstacles through vegetation at depths of 10 cm to several meters depending on density and vegetation distribu- tion [21]. Lalonde et al. [22] extract local point statistics such as the amount of scattering, linearity and plane approximations to train a gaussian mixture model. The model shows im- pressive results for forested environments, separating ground, tree trunks and canopies. Doerr et al. [23] compare methods for detecting foreign objects hidden in crops. The methods include tests for connectivity, discontuity, average height and average density of adjacent points. In [24]
Wellington attempts to estimate a ground model based on dual-return and classifies obstacles based on the ground model. Hollinger et al. [25, 26] attempt to exploit dual-returns from a commercial LiDAR sensor for obstacle detection for road-side hazards. They found that highly occluded targets are challenging to detect.
LiDARs offer the capability of detecting partly occluded obstacles. The work of Lalonde et al [22] even indicate the ability to detect and classify highly occluded obstacles. These LiDAR systems are very high resolution systems that use a long time to acquire a full image and are therefore to some degree unsuitable for moving platforms. The results presented using systems suitable for moving platforms are not satisfactory. In addition LiDAR sensors are not able to penetrate vegetation.
2.2 Radar systems
2.2 Radar systems
There have been several attempts to detect obstacles through vegetation using radar, some of them will be reviewed here. Matthies et al. [21] experimented with an impulse-radar with a center frequency of 2 GHz and a bandwidth of 3 GHz. Using a rail and imaging tomography techniques they were able to detect a large rock hidden behind 2.5 m of thick foliage.
Johnsten [27] used a 77 GHz scanning fan-beam radar to aid obstacle detection during the 2005 DARPA Grand Challenge. A scanning fan-beam radar is system that scans with a rotat- ing antenna, with a narrow beamwidth in azimuth. He concludes that the radar is a valuable sensor aiding LiDAR in obstacle detection, with no mention of vegetation. Peynot et al. [28]
compare the capabilities of several perception sensors for autonomous vehicles, concluding that radar and/or LiDAR is necessary to be able to navigate challenging conditions like fog, smoke and heavy rain. Using the same platform Reina et al. [29] attempt ground segmentation for autonomous vehicles, using (only) a 24 GHz scanning fan-beam radar and their results indicate that radar-only navigation is possible.
One simple approach, with regards to hardware complexity, is presented by Ahtiainen [30, 31].
The system uses a combination of LiDAR sensor and a single radar with a center frequency of 6.35 GHz and a bandwidth of 400 MHz. The solution requires the platform to effectively scan the area, as they do not have an array, nor a scanning radar platform. Traversability maps are created using LiDAR, then the radar is used to update these maps to detect non-solid obstacles.
The solution is mainly a mapping solution, as it does not run online, however they claim an online solution is possible as well. The platform is also small and the maximum range of the radar is approximately 10 meters.
There have been some attempts at using commercially available Commercial Off-The-Shelf (COTS) components. Hollinger et al. [26, 32] utilize a Velodyne HDL-64E LiDAR, 24 GHz and 76 GHz radars in attempt to detect partially obscured roadside hazards. They found that many possible hazards like a paint can are not detectable using automotive COTS sensors.
Although not exactly the same application, U.S. Army Research Laboratory(ARL) has de- veloped a forward-looking ground-penetrating radar (FLGPR) for IED-detection. They have also investigated the use of a system for foliage-penetrating obstacle detection [33]. The devel- opment is explained in several articles, starting with an impulse radar [34–41] and continuing with a stepped-frequency solution [42–45]. The stepped-frequency solution uses a receive array of 16 antennas and two transmit antennas. A frequency range from 300 MHz to 2 GHz gives the system great penetration abilities, combined with high range resolution. The system is capable of detecting surface and shallow-buried land mines, in addition to surface obstacles.
The system closest to our application is developed by the Agency for Defense Development (ADD), republic of Korea. This system development is lead by Sun-Gu Sun [46–51] and in- spired by the work in [35]. This system uses a FMCW MIDO (Multiple Input, Dual Output) configuration, with a frequency range from 3.6 to 5.2 GHz. The receive antenna is a 16 chan- nel antenna-setup that spans the width of the vehicle. Their experiments show very promising results, demonstrating the ability do detect obstacles through vegetation in single radar images.
The ability to discriminate between vegetation and obstacles is debatable.
Radar systems show promising results, all be it lower resolution than contending sensor techno- logies. The tests done by Matthies et al. [21] point to the possibility of building a system. The system presented by Sun-Gu Sun [46] and Nguyen [45] look promising and are great starting points for designing a new system.
2.3 Vegetation Attenuation
The similar systems presented in the covered literature vary a lot in center frequency. Matthies et al. experiments with a frequency content from 1-4 GHz and a center of approximatly 2.2 GHz. Ahtiainen presents a solution with a center frequency of 6.35 GHz. Sun-Gu Sun [46] has a center frequency of 4.4 GHz and Nguyen [45] has a center frequency of approximately 1.2 GHz. Normal vehicle-mounted radars usually operate in the 24 or 76 GHz are. With such a wide spread of center frequencies in the available literature it is beneficial to conduct a study of how the attenuation through vegetation is affected by the center frequency of the radar system.
Both the frequency and geometry affect the propagation through foliage. Results from the literature review are quite conclusive with regards to frequency; lower frequencies allow for higher foliage penetration. Attenuation in foliage generally increases greatly with frequency.
In [52] the attenuation is reported to be 30.4, 12.7 and 5.4 dB/m at center frequencies of 5.3, 1.25 and 0.44 GHz respectively. All of these measurements were at a 45 degree depression angle. The absolute values vary greatly in the research due to variables like water content, type of vegetation etc, but the trend is replicated in other research [53–55]. These studies are all based on measurements from airborne platforms and/or with forests as their main challenge.
Little research has been conducted with regards to random small-grown foliage, but a lot of work has been conducted on structured agricultural vegetation. With Ulaby in the lead, sev- eral measurment campains and models are presented [1, 56–59]. These results confirm that the theory for airborne applications is indeed transferrable to small-grown foliage, with attenuation generally increasing with frequency. Attenuation for the respective frequencies is summarized in Table 2.1. As previously mentioned, the results confirm that the attenuation at lower bands is close to negligible, but it drastically increases with frequency. This studies also underline some important observations with regards to polarization; the attenuation increase in vertical polarization is generally dominated by the vertical stalks in the vegetation itself [1]. Indicating that the difference in attenuation might not be as great for more herbaceous vegetation.
This is consistent with similar findings for forested environments, where it is evident that tree trunks are the major contribution for attenuation for vertically polarized waveforms. Backs- catter from vegetation also has an increasing tendency with frequency, however, it is not as pronounced as the attenuation.
The total attenuation also increases with increasing depression angle, because of the increasing amount of vegetation. Backscattering has the opposite effect, it increases with increasing de- pression angle. Most likely due to the fact that the “wall” created by the stalks is closer to a specular angle.
As mentioned, there are presented very few results from similar platforms to the UGV in ques- tion. Ahtiainen presents a solution with a center frequency of 6.35 GHz with an attenuation variation of 5-20 dB with approximatly 10-40 cm of vegetation, a simple inversion would lead
2.4 Radar imaging systems and algorithms
us to believe this is approximately 50 dB/m, which is significantly higher than the other results.
Matthies et al. experiments with a frequency content from 1-4 GHz and a center of approximatly 2.2 GHz. There is no mention of any specific attenuation measurements, but from the results it is fair to conclude that the attenuation is acceptable. Nam and Sun [60] concludes that the attenuation in C-band is twice that of S-band with the same geometry, without any reference to precise frequencies. These results are hardly conclusive and indicate that further measurements would be beneficial.
Table 2.1: Summarization of the attenuation in corn-fields with experiments conducted by Ulaby et al. [1]
Frequency Attenuation vertical Attenuation horizontal [GHz] polarization [dB/m] polarization [dB/m]
1.6 8.57 2.82
4.75 10.59 10.63
10.2 15.37 16.29
2.4 Radar imaging systems and algorithms
This section will present some of the literature covering general radar mapping and imaging.
Radar imaging is done on a large scale every day through automotive phased array radars, providing vehicle and hazard detection in diverse weather conditions. Being a large contribu- tion the industry, these systems are not well represented in academia. The dominating radar configuration in robotic research is the mechanically scanning fan-beam configuration with a small beamwidth. This configuration has been used in the DARPA Grand Challenge [27]. The same platform has been used by Mullane [61]. A similar configuration has been utilized in the QUAD-AV project [62].
Imaging radar arrays are machanically simpler by construction, with no moving parts. They are however, more complex with regards to both imaging process and hardware. Conventional phased-arrays can be used for this purpose, but they require half wavelength between each element and imaging systems are often both large and expensive. Better performance can be achieved using a Multiple-Input and Multiple-Output (MIMO) configuration, increasing the resolution. The signal orthogonality can be achieved by several means, the most popular being Time-Division Multiplexing (TDM). Several promising and well functioning imaging systems have been demonstrated. Wilden presents the “MIRA-CLE X” [63] with 16 transmit and 14 receive antennas. Similar demonstrator systems have been presented by Ganis [64] and Char- vat [65]. The two systems presented by Nguyen [45] and Sun [48] are also of this configuration.
The imaging algorithms used in these publications vary. The “MIRA-CLE X” [63] uses both backprojection and conventional beamforming with results presented by Klare [66], whilst Charvat [65] uses the so-called Range Migration Algorithm (RMA). From the results presen- ted in several of these publications, time domain backprojection seems to be the best and most agile processing algorithm of the candidates. A thorough review of this algorithm is given by Duerch [67], while other methods are presented in the classical work by Carrara [68]. To formulate the time domain backprojection for a general acMIMO Bistatic configuration, the considerations presented by Willis [69] are helpful.
2.5 Summary
Alternative sensors, particularly LiDAR proves to be promising with respect to detection of partly occluded obstacles. Due to the high resolution and accuracy of the sensors, the clas- sification is impeccable. The methods presented by Lalonde [22] are impressive and worth improving, albeit not as part of this work. The results with LiDAR sensors suitable for usage on a moving platform with real time date processing are less successful. Daylight cameras are not suitable for detecting hidden obstacles, but their resolution capabilities make them exceptional at general object detection and situational awareness in good atmospheric and daylight condi- tions. IR and spectral cameras are to some degree able to fill this void even in more challenging conditions; low light, smoke and fog. However, they do unfortunately not present any good res- ults with regard to occluded obstacles, other than differentiating between man-made and natural objects.
Radar imaging systems use a variety of different configurations. The most promising archi- tecture for forward imaging is a MIMO array, combining both high resolution and low system complexity and cost. Several working demonstrators using this architecture have been presen- ted; in particular the “MIRA-CLE X” [63] and the systems presented by Charvat [65] seem viable. With regard to imaging algorithms, time domain backprojection appears to be concep- tually elegant and agile. Radar systems for robotic mapping have several configurations. The most common are beamforming arrays, fan-beam scanning radars and MIMO radar arrays. To be able to use the radar for more conventional radar tasks such as UAV detection, an array configuration is definitively beneficial. The MIMO array architectures are also interesting, as they achieve high imaging performance combined with relative low-complexity systems with dramatically lower cost. To be able to map/image the area some sort of imaging algorithm need to be employed. Conventional FFT based methods fall short as many of their prerequisites are not met in a short range imaging scenario, such as the far-field and plane wave approximations.
The main candidates remaining are the Range Migration Algorithm (RMA) and Time Domain Backprojection. Both are worth considering.
Radar systems definitely have the ability to detect hidden obstacles and penetrate a certain amount of matter, proven by ground penetrating radars and through-the-wall imaging systems.
Radar systems look like a viable option for imaging with the capability of detecting hidden obstacles. In the literature, some promising systems are presented. Particularly the systems presented by Nguyen [70] and by Sun [46] are promising. The systems have similar configura- tions and the underlying MIMO architecture, but drastically different center frequencies. Both systems are also UWB systems. The articles also fail to present mapped imagery of complex natural scenes.
The large variations in attenuation measurements are understandable due to local variations, hence they motivate additional measurements as to the attenuation relevant scenes. There is no doubt that lower frequencies result in lower attenuation, to maximize the resolution, it is beneficial to build as a system with the highest frequency possible. Using a MIMO imaging configuration appears to be beneficial, but needs a foundation of analytical work and simula- tions to quantify the performance increase.
Chapter 3
Radar introduction
The wordradaroriginates as an acronym for radio detection and ranging, fitting it’s very first applications detection of ships and airplanes. This is still one of the main applications of radar, but radar systems have evolved and are today used in many applications, a few being: weather radar, ground-penetrating radar, police radar and SAR. The systems used for these tasks vary immensely. From large long-range detection radars to small and short range medical radars.
The underlying principle is the same.
The goal of this chapter is to give a short and concise introduction the concept of radar. Covering both the general mathematics for signal modeling and the conservation of energy. The chapter also gives a quick explanation of some of the measurement equipment used in the thesis.
3.1 The concept of radar
A general radar works by transmitting Electromagnetic (EM) and receiving the signal reflected from an object. The transmitted waves propagate through the air or vacuum. When the EM waves reach a target, the energy is generally both reflected, transmitted and absorbed in the target. The EM waves induce a current on the surface of the target, these currents induces and gives rise to a reflected wave. A small portion of this energy will travel in the direction of the receiving antenna and will be the registered reflection. Electromagnetic waves travel at the speed of light through air and vacuum. The rangeR to a given target can be determined by the time it takes the EM wave to travel to the target and back:
R= ct0
2 (3.1)
Where cis the speed of light and t0 is the round trip time from the antenna to the target and back. An important part of radar is energy conservation. EM waves travel radially from the source to the target, reflects off and travel radially on the way back to the detector. Therefore only a small portion of the transmitted power hits the target and an even smaller portion hits the detector. The received echo will therefore be orders of magnitude lower than the transmitted signal. This relation is described by the radar range equation, equation 5.29a in [3]:
Pr = PtGtGρλ2
(4π)3R2tR2rσt (3.2)
wherePris the received echo power as function of the transmitted powerPt, Antenna gain of transmitting and receiving antennasGtandGr, the wavelengthλ, the radar cross section(RCS) σtof the target and round trip range to the target, given by the range to the transmitting antenna Rtand the receiving antennaRr. This equation states the transmit power needed to detect a tar- get of certain Radar Cross Section (RCS) at a certain distance. This equation can be rewritten to include system noise and represent the Signal to Noise Ratio (SNR). In real life the response rarely originates from one target, its a function of all the scatterers within a range cell. The signal from the target will contain interference. Interference can generally originate from three sources: internal and external electronic noise, reflections off object of no interest (clutter) and interference from other spectrum users.
3.2 Radar Waveforms
Radars can be divided into two general categories: Continuous Wave (CW) and pulsed. The CW waveform continuously transmits and receives the signal. Whereas the pulsed configuration cannot transmit and receive at the same time. This type of radar transmits a pulse before switch- ing to receive. Both categories have their strengths and weaknesses. Pulsed configurations have a “blind zone” that corresponds to the distance the wave has traveled during the transmission of the wave, comparable to the pulse length. CW radars often struggle with isolation between transmitter and receiver, therefore they experience signal “leakage” from transmitter to receiver.
This thesis will only cover CW radar in detail.
3.2 Radar Waveforms
3.2.1 CW Waveform
The CW waveform is a pure sinusoid that is being continuously transmitted. The instantaneous frequency is equal to:
f(t) = f0 (3.3)
Giving us a phase of:
φ(t) = 2π Z t
0
f(τ)dτ = 2πf0t (3.4)
Leaving us with a transmitted signal of:
Stx =Aej2πf0t (3.5)
If we picture a scene with one point scatterer at a location corresponding to the time delay t0, the received signal will be a time-delayed copy of the transmitted signal:
Srx =Kej2πf0(t−t0) (3.6)
The phase of the received signal can be determined by comparing the received signal with the transmitted signal. This process is called homodyning and is in practice a multiplication of the two signals.
Sif = AK
2 ej2πf0tej2πf0(t−t0) (3.7)
= AK
2 ej2πf0tej2πf0te−j2πf0t0 (3.8)
= AK
2 ej2π2f0t−j2πf0t0 (3.9)
Leaving us with one signal at double frequency and one signal at baseband. Ais the amplitude of the transmitted signal and K can be calculated using the radar equation. Using a low-pass filter, we can obtain the Intermediate Frequency (IF):
Sif = AK
2 e−j2πf0t0 (3.10)
where
t0 = 2R
c (3.11)
which means that the complete phase is equal to:
φ =ωt0 = 2πf2R
c = 4πR
λ (3.12)
This is a real signal, so it can also be expressed as:
Sif = AK
2 cos 4πR λ
!
(3.13) As we can see from Equation 3.13, the output voltage will be very sensitive to changes in range between the radar and target. Unfortunately we do not have any unambiguous information about the actual range to the target.
3.2.2 FMCW waveform
As mentioned in Section 3.2.1, a pure CW radar has no useful range-information. This is expected as it in practice transmits a “pulse” of infinite length. If we add a frequency modulation to the transmitted signal, we can extract range information as well.
The most common type of frequency modulation is Linear Frequency Modulation (LFM). The frequency of the transmitted signal will modulate linearly fromf0 tof1 with a center atfc. The bandwidth of the transmitted signal will beB =f1−f0. This type of pulse is sometimes called achirp. In the time domain, the transmitted signal can look like Figure 3.1.
Figure 3.1:Time domain representation of a LFM chirp
The instantaneous frequency is equal to:
f(t) =fc+γt (3.14)
Withtbeing the fast time variableT ∈ [−T < T], whereT is the pulse length.fcis the center frequency andγ is the chirp rate defined as the bandwidth divided by pulse lengthγ = BT. The corresponding phase is:
φ(t) = 2π Z t
0
f(τ)dτ (3.15)
= 2π Z t
0
(fc+γt) dt= 2π
fct+ γ 2t2
(3.16) whereφ0is the initial phase, which is removed for simplicity. Giving us a transmitted signal of:
Stx =Aej2πfctejπγt2 (3.17)
Given the same logic as with CW, the received signal will be an identical signal with a time delayt0 and a different amplitudeK:
Srx=Kej2πfc(t−t0)ejπγ(t−t0)2 (3.18)
3.2 Radar Waveforms Mixing these two together, gives us the IF signal:
Sif =SrxStx (3.19)
= KA
2 ej2πfctejπγt2ej2πfc(t−t0)ejπγ(t−t0)2 (3.20)
= KA
2 ej2πfctej2πγt2ej2πfctej2πfct0ej2πγt2ej2πγtt20e−j2πγt0ej2πγ2tt0 (3.21)
= KA
2 e−j2πfct0ej2πγtt0ej2πfct20ej2π(2fct+γt2) (3.22) The last of these terms is the double-frequency component. Using a low-pass filter leaves us with:
Sif = KA
2 e−j2πfct0ej2πγtt0ej2πfct20 (3.23) The components of this signal are named in Table 3.1 according to Carrara [68]. Theamplitude is dependent on the targets reflectivity. The Range dependent phase is responsible for posi- tioning our scatterer within each range cell. It is this term that later will be used in the SAR processing to generate azimuth resolution. Conversely the Complex range dependent sinusoid is responsible for positioning the scatterer in range. TheResidual video phase(RVP) is an un- wanted effect of the homodyning process and is in most cases very small [68], it is therefore not included in our simulations.
Table 3.1: Components of FMCW IF signal Component Expression
Amplitude KA2
Range dependent phase e−j2πfct0 Complex range dependant sinusoid ej2πfctt0 Residual video phase ej2πfct20
To give a conceptual explanation to how FMCW radars acquire range information Figure 3.2 illustrates the transmitted frequency and the response from a target. The response from a target is a time-delayed attenuated copy of the transmitted signal and theintermediate frequency,fI, is proportional to the rangeR, or more precisely the time delayt0. The transmitted signal form a triangle with legs of lengthBandT. The IF frequencyfIalso forms a rectangle with the time delayt0. These triangles are similar. We can therefore express their relationship:
fi t0 = B
T (3.24)
which leads to:
fi = 2γ
c Rt (3.25)
Whereγ once again is the chirp rate,cthe speed of light andRt the range to the target. From this equations, it is evident that by measuringfi, we can calculate the range to the target, given that both the bandwidthBand pulse lengthT are known.
Figure 3.2:Illustration of the FMCW principle, showing the response of a single ideal point scatterer.
3.3 Electromagnetic wave propagation
Electromagnetic waves are electric and magnetic field waves. The nature and relations of elec- tromagnetic fields are described by Maxwell’s equations. The electric field and magnetic field travel are perpendicular to each other, and also perpendicular to the direction of the traveling wave. The direction of the wave is orthogonal to the plane created by the electric and magnetic fields. The interaction between EM waves and materials involves scattering, absorption, trans- mission, and emission [3]. Most of the theory explaining these phenomenon use the definition of a plane wave. A wave radiated from a source has a spherical wavefront, it radiates in all dir- ections at the same speed. If an observer is sufficiently far away from this source, it is described as the far-field and the wavefront can be considered to be plane as illustrated in Figure 3.3.
Figure 3.3:Waves radiated from a source and approximated by a plane wave in the far-field.
3.3 Electromagnetic wave propagation
3.3.1 Constitutive parameters
The parameters of a material that control how EM waves interact with it are called the con- stitutive parameters [3];
ε0ε0 =electrical permittivity(F/m) (3.26)
µ=magnetic permeability(H/m) (3.27)
ρv =volume charge density(C/m3) (3.28)
σ=conductivity(S/m) (3.29)
The parameters affecting non-magnetic charge-free materials areε0ε0andσ. We can also define the complex dielectric constantεas:
ε=ε0−j σ
ωε0 =ε0−jε00 (3.30)
ε00 = σ
ωε0 (3.31)
Whereε0is the materials permittivity andε00is known as the loss factor. Using these parameters, we can model how waves travel through a medium and how they react to changes in material properties.
3.3.2 Polarization
In many radar applications, we use an active transmitter to illuminate the scene. This gives us the ability to control parameters of the transmitted wave. One attribute we would like to control is the polarization. In the most general case, the tip of the electric field vector forms an ellipse over time. However, in more specialized cases, it can degenerate to a circle or a straight line.
This is known as circular and linear polarization. Most radar applications use either circular or linear polarization. In the linear case, we also need to specify in which direction it forms a straight line, either vertical or horizontal, illustrated in Figure 3.4.
Figure 3.4:Illustration of the polarization of the electric and magnetic field. Adapted from Ulaby [3]
3.3.3 Reflection and transmission
When an electromagnetic wave travels from one medium to another, the wave will generally be both reflected from the medium and transmitted into the material. How much energy is reflected and transmitted is governed by the materials properties, described by its constitutive paramaters, and expressed through the Fresnel reflection and transmission coefficients. If the angle of incidence is oblique these effects also vary depending on the polarization of the incident wave.
3.4 Radar scattering
The concept of radar is based on scattering of EM waves, as illustrated in Figure 3.5. We have introduced the concepts of reflection and transmission. Reflection is analogous to what you see when you look in the mirror. The light bouncing off you, hits the mirror and get reflected back.
However, if you look at a brushed steel surface, you might be able to see the contour of yourself, but it is blurred - the light has been scattered. This phenomenon also happens with radio waves.
When EM waves hit a rough surface, they get scattered in multiple directions. The components of the scattering coming back to the radar is known as backscatter. The backscattered wave can be a result of surface scattering or volume scattering.
Figure 3.5: Illustration of the scattering of an incident beam by a target, some vector graphics provided by vecteezy.com.
If the boundary defining the intercepted radar signal is continuous and no layers below this boundary contributes to the backscattered signal, we can define it as surface scattering. The sur- face parameters defining the scattering is the dielectric constant as mentioned in Section3.3.1, surface roughness, height and height correlation. These parameters in turn react with the radar frequency, incidence angle and the polarization of the radar. From a surface, we will have direct reflections from the surface, as well as multi reflections as illustrated in Figure 3.6
Figure 3.6:Illustration of a multi-path reflection from a rough surface.
3.5 Foliage Phenomenon
Volume scattering occurs when the the defining boundary between two materials is not continu- ous. The waves enter the volume and scattering is caused by discrete particles present within that volume. A vegetation canopy is an example of a scattering volume, where the incident wave enters the volume and scattering is caused by the vegetation.
3.5 Foliage Phenomenon
For foliage penetration radars, foliage would contribute to volume scattering. In addition, there might be a certain amount of surface scattering from the ground below. The effects are the same as previously mentioned, but their terminology might differ. The effects that will be visible in an imaging radar system are summarized by Richards [71]: attenuation, backscatter, phase variation and depolarization.
• Attenuation happens when the EM waves propagates through vegetation, this effect is comprised of absorption and scattering.
• Backscatter is the part of the wave that gets transmitted back to the radar from the veget- ation itself, either in the surface or throughout the volume.
• Phase variation occurs because of the nonuniform distribution of the scatterers in vegeta- tion.
• Depolarization occurs for the same reason as phase variation, the incident wave is scattered multiple times throughout the volume, changing the polarization state.
All of these effects vary depending on the frequency, viewing angle and the denseness of the vegetation.
3.5.1 Foliage terminology
To describe the attenuation through foliage, we need to be able to describe the type of vegetation.
There is a short introduction to some of the structural elements of vegetation in Figure 3.7. Most of the existing research and measurements has been conducted on either trees or agricultural crops. Trees contain a combination of foliage, branches and trunks. Most agricultural crops contain blade-like herbaceous vegetation. The agricultural crops are probably most similar to our problem. From the pictures presented in Chapter 1, we can deduce that our problem contain mostly blade-like herbaceous vegetation, but it can also include broadleaf and shrubs.
Figure 3.7: Description of structural classes of vegetation as described by Dobson [4], some vector graphics provided by vecteezy.com.
3.5.2 Foliage parameters
There are many variables defining the foliage itself, in addition to the structural classification in Figure 3.7, the following parameters also greatly affects their properties:
• Volume fraction, the amount of vegetation relative to air.
• Distribution
• Shape
• Water content
• Surface properties
All off these variables vary throughout the lifetime of a given vegetation type, but it can also vary based on current and/or present weather conditions.
3.6 Radar equipment used in this thesis
All the measurements presented in this thesis were conducted using either Vector Network Analyzer (VNA), Mixed Signal Oscilloscopes or the Radar on an Off-road Light Autonomous Vehicle (ROLAV) radar prototype.
3.6.1 Vector Network Analyzer
The VNA used in most of the measurements in this thesis is the Rohde Schwarz ZVA 67 [72].
The VNA can transmit signals from 10 MHz to 67 GHz and is possible to remote control over a network. Generally VNAs are used to measure the amplitude and phase parameters of an electrical network. By connecting antennas, however, the scene becomes the measured
“network”. The VNA can record the so-called S-parameters of a network. The parameter Snm is the ration between the transmitted voltage on port m and the received voltage on port n. A directional coupler is connected to each port, enabling the measurement of S11 andS22 which relates to both the reflection and transmission through the network. The S-parameters are recorded as a function of frequency. A Python script was developed to remote control the VNA using the VNA instrument driver1to enable easy data retrieval.
3.6.2 Synthetic aperture radar rail
Synthetic Aperture Radar (SAR) imaging sensors are useful for measurements in controlled environments where the measured scene is fixed. In general, they provide the same performance as a large array based system using only one TX/RX pair. To be able to measure attenuation and target contrast, in addition to test resolution capabilities, a SAR rail imaging sensor was developed as a part of the thesis. The system uses a linear rail with a stepper motor, controlled by an Arduino micro controller via USB connection to the host computer. The components of the system are shown in Figure 3.8.
1Available at: https://github.com/Terrabits/rohdeschwarz
3.6 Radar equipment used in this thesis
Figure 3.8:High level diagram showing the components of the SAR rail. The host computer is connected via Ethernet to the VNA and USB to the Arduino micro controller. Using a python script on the host computer, the data is collected and the antenna positions recorded.
The antenna assembly can be mounted to the rail carriage, enabling a synthetic aperture of up to 180 cm. Using a stop-and-go approach the system homes itself, captures a sample and moves a pre-defined distance and repeats until it reaches the end of the rail. Acquiring a high-quality measurement with all the S-parameters using the VNA is time consuming, so a full rail scan takes about 30 minutes. Placing the rail itself on tripods, results in a semi-portable solution as shown in Figure 3.9 Using a python script to both remote control the rail and the network analyzer, enables easy automated recording of radar images. The full workflow of the system is shown in Figure 3.10.
Figure 3.9:Image showing the SAR rail setup
Start Initialize VNA Move to home
Move one
interval Acquire sample
yes
no Captured
desired samples?
Finished
Figure 3.10:Flowchart of the SAR rail operation
Coupled with broadband antennas, the systems performance is unparalleled enabling SAR ima- ging with a 180 cm aperture and 13 GHz bandwidth. To illustrate the systems performance, FFI was spelled on a styrofoam sheet using nails. This sheet was then imaged using the meas- urement setup resulting in the image presented in Figure 3.11. To acquire this image, coherent