Xie, Zhenming2023-02-032023-02-032022-11https://hdl.handle.net/11299/252319University of Minnesota Ph.D. dissertation. November 2022. Major: Mechanical Engineering. Advisor: Rajesh Rajamani. 1 computer file (PDF); viii,88 pages.Perception or scene understanding is one of the major problems in Advanced Driver Assistance Systems (ADAS), full autonomous driving, and anything in between. It is also a major component of many robotic systems and intelligent transportation systems. Object tracking is a common approach to understanding the dynamics of various moving objects and hazards in the scene. Vehicle tracking can be seen as a specific type of object tracking, where some vehicle-specific models, assumptions and methods can be used. These specifics provide bridges, some of which are explored in this dissertation, to enable and improve vehicle tracking even with limited sensor information. In this dissertation, vehicle tracking algorithms are developed for two types of sensors, low-density Lidar (light detection and ranging) sensors and millimeter-wave Radar (radio detection and ranging) sensors. Both types of sensors are relatively low-cost but provide relatively limited measurement information about target vehicles. Specifically, the two-dimensional low-density Lidar used in this work provides only several valid distance measurements from its 8 segments, each being 6 degrees wide, resulting in a large azimuth measurement uncertainty. While the millimeter-wave Radar sensor provides better azimuth resolution in its point cloud, measurements may vary significantly between frames depending on the reflection surface and possible micro-doppler effects. The objective of this work is developing algorithms that maximize vehicle tracking performance from the limited sensor measurement information. This also enables the possibility of new intelligent transportation applications with the aforementioned sensors. This dissertation mainly includes three parts with different algorithms and application scenarios, which are covered in three chapters. Each chapter includes algorithm development and evaluation with both simulation and real-world experiments. In chapter 2, a vehicle tracking algorithm for a low-density Lidar is developed, which forms the basis for a bicyclist protection system against collision with motor vehicles. The developed system detects if a nearby vehicle poses a collision danger to the ego bicycle and provides an audio warning to alert the motor vehicle driver of the imminent collision. The main challenge with the solid-state low-density Lidar is its low azimuth resolution. To handle this, an adaptive correction term is computed for each of the measurement points in each sampling to obtain a synthetic position measurement, which is near the closest point on a vehicle with respect to the ego bicycle. The system then tracks the approximate closest point on each vehicle and makes warning decisions based on the predicted trajectories of (the closest point of) each vehicle. A multi-vehicle tracking framework is developed for handling scenes with multiple vehicles. Hierarchical clustering with a custom distance measure is used for detecting vehicles from raw measurement points. Data association is handled by the global nearest neighbor method. Vehicle kinematic states are estimated by interacting-multiple-model extended Kalman filtering. Good multi-vehicle tracking results are achieved despite the low sensor resolution. Chapter 3 extends the work of chapter 2 to develop a vehicle counting and maneuver classification algorithm with the same low-density Lidar sensor for stationary road-side traffic monitoring systems. Vehicle tracking is performed similarly to chapter 2, but with an unscented Kalman filter for improved filtering performance. Then, characteristic low-dimensional features of each vehicle trajectory are extracted for maneuver classification with support vector machines. These features include selected vehicle locations and velocities, and singular vectors that describe the shape of the trajectory around the mean vehicle location. These features provide excellent separations between various inlet-outlet maneuvers at a traffic intersection. Both a single-sensor setup and a two-sensor setup are considered. The single sensor setup that covers 2 out of 4 roads at a traffic intersection is found to work with high accuracy but has occlusion errors due to the limited coverage of the sensor. The two-sensor setup that covers all 4 roads at an intersection needs an additional algorithm to merge trajectories from the two sensors to avoid double counting. High counting and classification accuracies are achieved with the developed systems. Chapter 4 proposes a vehicle tracking framework with Radar measurements using a novel unscented Kalman filter based on the maximum correntropy criterion (MCC-UKF). A single Radar sensor still provides limited sensor information and may include false alarms in the measurements. But compared with the low-density Lidar used in previous chapters, the millimeter-wave Radar sensor used provides a better azimuth resolution in its point cloud, and doppler range rate measurements. These enable the possibility of using more sophisticated models and methods that maximize the use of Radar sensor information for better vehicle tracking performance. A nonlinear kinematic vehicle model incorporating vehicle dimension estimation is proposed as the system dynamic model. Two measurement models for millimeter-wave Radars are used alternatively depending on measurement conditions. Data association is efficiently handled by assigning individual measurement points to established vehicle estimates based on their statistical distances. Two-dimensional vehicle dimensions are estimated alongside vehicle kinematics to aid data association. A two-stage filtering scheme is used for robust initialization and tracking. An adaptive kernel width based on the estimated state covariance is used in the MCC-UKF for noise/clutter rejection during vehicle tracking. The MCC-UKF-based method’s ability of rejecting non-Gaussian noise propels the use of more sophisticated models developed in this work, which can otherwise be sensitive to clutters. The proposed framework is relatively computationally lightweight for real-time implementations and shows overall improved vehicle tracking performance when compared to baseline methods in both simulations and real-world experiments.enThesis or Dissertation