LiDAR Based Perception System: Pioneer Technology for Safety Driving

Abstract

Perceiving the surrounding multiple vehicles robustly and effectively is a very important step in building Advanced Driving Assistant System (ADAS) or autonomous vehicles. This thesis presents the design of the Light Detection and Ranging (LiDAR) perception system which consists of several sub-tasks: ground detection, object detection, object classification, and object tracking. It is believed that accomplishing these sub-tasks will provide a guideline to a vast range of potential autonomous vehicles applications. More specifically, a probability occupancy grid map based approach was developed for ground detection to address the issues of over-segmentation, under-segmentation and slow-segmentation by non-flat surface. Given the non-ground points, point cloud clustering algorithm is developed for object detection by using a Radially Bounded Nearest Neighbor (RBNN) method on the static Kd-tree. To identify the object, a supervised learning approach based on our LiDAR sensor for vehicle type classification is proposed. The proposed classification algorithm is used to classify the object into four different types: ``Sedan'', ``SUV'', ``Van'', and ``Truck''. To handle disturbances and motion uncertainties, a generalized form of Smooth Variable Structure Filter (SVSF) integrated with a combination of Hungarian algorithm (HA) and Probability Data Association Filter (PDAF), referred to as GSVSF-HA/PDAF, is developed. The developed approach is to overcome the multiple targets data association in the content of dynamics environment where the distribution of data is unpredictable. Last but not the least, a comprehensive experimental evaluation for each sub-task is presented to validate the robustness and effectiveness of our developed perception system.