Enhancing Perception for Autonomous Vehicles
Autonomous Vehicle , Deep Learning , Computer Vision , Object Detection , Semantic Segmentation , Cooperative Perception
Autonomous Vehicles (AVs) have made great progress with the advancements in high performance computing and Artificial Intelligence (AI) in recent years. AVs are equipped with Automated Driving Systems (AVS) that are able to manipulate the environment and perform driving tasks safely without human intervention. With a precise perception, AVs can analyze the traffic scene, localize the traffic participants, and then predict their motions to maneuver through traffic on the road. However, there is still hesitation about embracing the technology and skepticism about the reliability and robustness of the ADS in the fickle and noisy traffic environment. Current perception systems equipped with RADAR, camera and LiDAR still face great challenges caused by occlusion, resolution, and weather condition. In this research, we focus on vehicular data analysis using camera and LiDAR data, and apply deep learning-based model frameworks to solve and improve multiple AV perception tasks including dynamic traffic participants detection and road segmentation. We also explore cooperative perception among Connected Autonomous Vehicles (CAVs) with the Vehicular Communication (VC) systems to improve the perception of distance and accuracy. Cooperative perception allows a CAV to interact with the other CAVs in the vicinity to enhance perception of surrounding objects as well as increase the safety and reliability of AVs. It can compensate for the limitations of the conventional vehicular perception such as occlusion, blind spots, low resolution, and weather effects. This thesis presents our work with regard to enhancing perception of AVs including camera-based vehicle detection, camera-based road segmentation and drivable area detection, LiDAR-based 3D object detection, LiDAR-camera fusion-based 3D object detection and Bird's-Eye View (BEV) semantic segmentation. The experiments demonstrate that utilizing the cooperative perception outperforms the conventional single vehicle perception approaches.