Autonomous vehicles are complex robotic and artificial intelligence systems working together to achieve safe operation in unstructured environments. The objective of this work is to provide a foundation to develop more advanced algorithms for off-road autonomy. The project explores the point cloud data captured from lidar sensors, and the processing to restore some of the geometric information lost during sensor sampling. Because ground truth values are needed for quantitative comparison, the MAVS was leveraged to generate a large off-road dataset in a variety of ecosystems. The results demonstrate data capture from the sensor suite and successful reconstruction of the selected geometric information. Using this geometric information, the point cloud data is more accurately segmented using the SqueezeSeg network.
Proc. SPIE. 10658, Compressive Sensing VII: From Diverse Modalities to Big Data Analytics
KEYWORDS: Radar, Signal to noise ratio, Detection and tracking algorithms, Sensors, Matrices, Error analysis, Estimation theory, Receivers, Analog electronics, Motion models, Radar signal processing, Statistical signal processing, Compressed sensing
Compressive sensing theory states that a sparse vector x in dictionary A can be recovered from measurements y = WAx. For recovery of x, the measurement matrix W is generally chosen as random since a random W is sufficiently incoherent with a given basis A with high probability. Although Gaussian or Bernoulli random measurement matrices satisfy recovery requirements, they do not necessarily yield the best performance in terms of minimal mutual coherence or best parameter estimation. In literature several studies focused on measurement matrix design mainly to minimize some form of coherence between W and A to minimize measurement numbers while exact reconstruction is guaranteed. On the other hand, for enhanced parameter estimation W can be designed to minimize the Cramer Rao Lower Bound (CRLB). In this study, we propose direct and sequential measurement designs that minimizes the CRLB for the application of direction of arrival (DoA) estimation. Based on our results an adaptive target tracking procedure for single and multiple target scenarios is also proposed. Initial simulations show that measurement design solutions provide enhanced parameter estimation and target tracking performance compared to widely used random matrices in compressive sensing.
Many application areas including signal and image processing, computer vision, radar and remote sensing, bioinformatics deal with high dimensional data of various types. In these applications, the high dimensional data is not generally distributed over the whole signal space; rather it lives in the union of low dimensional subspaces. Hence, classical clustering techniques depending data distributions in centroids are not successful, and techniques that facilities the low dimensional subspace structure of big data are required. Sparse subspace clustering (SSC) technique that relies on the self-expressiveness of the data is shown to provably handle the data under noiseless case for independent and disjoint subspaces. Self-expressiveness means that each data point in a union of subspaces can be efficiently represented as a linear or affine combination of data points in the set. SSC implementation involves solving an L1 minimization problem for each data point in the space and applying spectral clustering to the affinity matrix constructed by the obtained coefficients. Despite good properties, SSC suffers from high computational complexity increasing with data point numbers. In addition, for noisy data self-expressiveness does not apply anymore. This paper proposes to use perturbed orthogonal matching pursuit (POMP) within SSC framework for robust and computationally efficient estimation of the number of subspaces, their dimensions, and the segmentation of the data into each subspace. POMP was shown to be successful in recovering sparse signals under random basis perturbations, which is actually the case in corrupted data clustering. Our initial results for simulated clustering datasets show that the proposed POMP- SSC technique provides both computational efficiency and high clustering performance compared to classical SSC implementation.
Cognitive radar is a novel concept for next-generation radar systems, which as part of the perception-action cycle to improve the measurement process based on dynamic changes in the environment. Although most work in this area to-date have focused on adaptation on the transmitted waveform, in this paper, we propose adaptive control of novel multifunctional reconfigurable antennas (MRAs) as a mechanism for action within the cognitive radar framework. Reconfigurable parasitic layer based MRAs have the capability of dynamically and simultaneously changing its electromagnetic characteristics (mode of operation), e.g. antenna beam pattern, polarization, center frequency, or a combination of thereof. Different modes of an MRA are controlled via RF switches interconnecting the pixels of the reconfigurable parasitic layer. This enhanced capability can be controlled using adaptive mode selection schemes. In particular, an array of MRAs provides more degrees of freedom, where each element of an array can be controlled to generate one of many modes depending on the environmental measured variables as a feedback mechanism. In this work, a designed and fabricated reconfigurable parasitic layer based MRA operating over 4.94-4.99 GHz band with 25 different radiation patterns, i.e., modes of operation, is utilized for cognitive direction-of-arrival (DoA) estimation and target tracking. A novel computationally efficient iterative mode selection (IMS) technique for MRA arrays is developed, where the modes are cognitively selected to minimize the DoA estimation error in target track. It is demonstrated that the proposed cognitive mode selection for MRA arrays achieves remarkably lower estimation errors compared to uniform pattern arrays without adaptive capability.
The human micro-Doppler signature is a unique signature caused by the time-varying motion of each point on the human
body, which can be used to discriminate humans from other targets exhibiting micro-Doppler, such as vehicles, tanks,
helicopters, and even other animals. Classification of targets based on micro-Doppler generally involves joint timefrequency
analysis of the radar return coupled with extraction of features that may be used to identify the target.
Although many techniques have been investigated, including artificial neural networks and support vector machines,
almost all suffer a drastic drop in classification performance as the aspect angle of human motion relative to the radar
increases. This paper focuses on the use of radar networks to obtain multi-aspect angle data and thereby ameliorate the
dependence of classification performance on aspect angle. Knowledge of human walking kinematics is exploited to
generate a fuse spectrogram that incorporates estimates of model parameters obtained from each radar in the network. It
is shown that the fused spectrogram better approximates the truly underlying motion of the target observed as compared
with spectrograms generated from individual nodes.
KEYWORDS: General packet radio service, Data modeling, Associative arrays, Ground penetrating radar, Radar imaging, Signal to noise ratio, Detection and tracking algorithms, Data acquisition, Imaging systems, Target detection
Spatial sparsity of the target space in subsurface or through-the-wall imaging applications has been successfully used within the compressive-sensing framework to decrease the data acquisition load in practical systems, while also generating high-resolution images. The developed techniques in this area mainly discretize the continuous target space into grid points and generate a dictionary of model data that is used in image-reconstructing optimization problems. However, for targets that do not coincide with the computation grid, imaging performance degrades considerably. This phenomenon is known as the off-grid problem. This paper presents a novel sparse ground-penetrating radar imaging method that is robust for off-grid targets. The proposed technique is an iterative orthogonal matching pursuit-based method that uses gradient-based steepest ascent-type iterations to locate the off-grid target. Simulations show that robust results with much smaller reconstruction errors are obtained for multiple off-grid targets compared to standard sparse reconstruction techniques.
A broadband quadrapole electromagnetic induction (EMI) array with one transmitter and three receiver coils
is built for detecting buried metallic targets. In this paper, it is shown that the locations of multiple metallic
targets including their depth and cross-range position can be estimated accurately with the EMI array using
an orthogonal matching pursuit (OMP) approach. Conventional OMP approaches use measurement dictionaries
generated for each possible target space point which results in huge dictionaries for the 3D location problem.
This paper exploits the inherent shifting properties of the scanning system to reduce the size of the dictionary
used in OMP and to lower the computation cost for possibly a real-time EMI location estimation system. The
method is tested on both simulated and experimental data collected over metal spheres at different depths and
accurate location estimates were obtained. This method allows EMI to be used as a pre-screener and results in
valuable location estimates that could be used by a multi-modal GPR or other sensor for enhanced operation.
Multimodal detection of subsurface targets such as tunnels, pipes, reinforcement bars, and structures has been
investigated using both ground-penetrating radar (GPR) and seismic sensors with signal processing techniques
to enhance localization capabilities. Both systems have been tested in bi-static configurations but the GPR has
been expanded to a multi-static configuration for improved performance. The use of two compatible sensors
that sense different phenomena (GPR detects changes in electrical properties while the seismic system measures
mechanical properties) increases the overall system's effectiveness in a wider range of soils and conditions. Two
experimental scenarios have been investigated in a laboratory model with nearly homogeneous sand. Images
formed from the raw data have been enhanced using beamforming inversion techniques and Hough Transform
techniques to specifically address the detection of linear targets. The processed data clearly indicate the locations
of the buried targets of various sizes at a range of depths.
Multi-static ground-penetrating radar (GPR) uses an array of antennas to conduct a number of bistatic operations simultaneously. The multi-static GPR is used to obtain more information on the target of interest using angular diversity. An entirely computer controlled, multi-static GPR consisting of a linear array of six resistively-loaded vee dipoles (RVDs), a network analyzer, and a microwave switch matrix was developed to investigate the potential of multi-static inversion algorithms. The performance of a multi-static inversion algorithm is evaluated for targets buried in clean sand, targets buried under the ground covered by rocks, and targets held above the ground (in the air) using styrofoam supports. A synthetic-aperture, multi-static, time-domain GPR imaging algorithm is extended from conventional mono-static back-projection techniques and used to process the data. Good results are obtained for the clean surface and air targets; however, for targets buried under rocks, only the deeply buried targets could be accurately detected and located.
We describe an efficient approach for finding probable target areas quickly with a minimal number of Ground Penetrating Radar (GPR) measurements. Since a potential GPR target creates a hyperbolic signature in the space-time domain, our approach uses the time delay differences from consecutive GPR A-Scan data to estimate the location of the apex of the hyperbolic signature, thus locating a target. This apex prediction method uses many fewer measurements than a full backprojection algorithm. Regions of low target probability are determined using a Neyman-Pearson detection approach in order to eliminate redundant measurements. In this regard, our approach is especially suitable as a pre-screener: other sensors that are more accurate, but require more measurement time, can then be applied only to high probability-of-target areas to corroborate results, differentiate between targets, or provide more accurate location measurements. Compared to a standard backprojection algorithm more signal-to-noise ratio (SNR) is needed to achieve similar detection performance. This SNR loss can be reduced by using a more conservative algorithm which reduces the step size of the GPR antenna. Results from experimental data collected at a model mine field at the Georgia Institute of Technology show that target positions can be found accurately using less than 10% of the measurements utilized by conventional imaging algorithms.
The imaging of subsurface targets, such as landmines, using Ground
Penetrating Radar (GPR) is becoming an increasingly important area
of research. Conventional image formation techniques expend large
amounts of computational resources on fully resolving a region,
even if there is a large amount of clutter. For example, standard
backprojection algorithms require O(N3). However, by using
multi-resolution techniques-such as quadtree-potential targets and clutter can be discriminated more efficiently with O(N2log2N). Because prior work has focused on the imaging of surface targets, quadtree techniques have mostly been developed for 2D imaging. Target depth adds another dimension to the imaging problem; therefore, we have developed a 3D quadtree algorithm. In this case, the mine field is modeled as a volume that is sub-divided at each stage of the quadtree algorithm. From each of these sub-volumes, the energy intensity is calculated. As the algorithm proceeds to finer resolutions, the energy in region containing a potential target increases, while that of background noise decreases. A multi-stage detector applied on intermediate quadtree data uses this change in energy to discriminate between regions of targets and clutter. This is advantageous because only the regions containing likely targets are investigated by additional sensors that are relatively slow in comparison to GPR (e.g. seismic or EMI sensors). This algorithm is tested on synthetic and experimental data collected from a model mine field at Georgia Institute of Technology. Even under near field and small aperture conditions, which hold for the mine detection case, test results show that target location information can be gathered with processing using the 3D quadtree algorithm.
A multi-static ground-penetrating radar (GPR) has been developed to investigate the potential of multi-static inversion algorithms. The GPR consists of a linear array of six resistively-loaded vee dipoles (RVDs), a network analyzer, and a microwave switch matrix all under computer control. The antennas in the array are spaced 12cm apart so the spacing between the transmitter and the receiver pairs in the measurements are from 12cm to 96cm in 12cm increments. The size of the array is suitable for the landmine problem and scaled measurements of the buried structure problem. The RVD is chosen as an array element because it is very "clean" in that it has very little self clutter and a very low radar cross section to lessen the reflections between the ground and the antenna. The shape and the loading profile of the antenna are designed to decrease the reflection at the drive point of the antenna while increasing the forward gain. The antenna and balun are made in a module, which is mechanically reliable without significant performance degradation. The multi-static GPR operation is demonstrated on targets buried in clean sand and targets buried under the ground covered by rocks. The responses of the targets are measured by each transmitter-receiver pair. A synthetic aperture, multi-static GPR imaging algorithm is extended from conventional monostatic back-projection techniques and used to process the data. Initial images obtained from the multi-static data are clearer than those obtained from bistatic data.