Towards Efficient Point-Cloud Object Detection on Autonomous Vehicles

Abstract

Autonomous vehicles (AVs) must perceive and understand the 3D environment around them. Modern autonomous vehicles use a suite of sensors and a set of machine-learning-based recognition algorithms to try to accomplish this goal. While such a system can achieve high accuracy, its cost is also prohibitively high for many applications. In this thesis, we aim to make the perception system of an AV both less expensive and more accurate. We design a multi-pronged approach that addresses challenges ranging from data representation, sensor configuration and training efficiency to the semantic understanding of road scenarios. We start with the data presentation of lidar point clouds, a primary input to many perception systems, and design a novel graph representation of point clouds to detect objects. Secondly, we study a sensor configuration of sparse lidars and complementary low-cost cameras. We propose a fusion method to combine video streams with sparse lidar point clouds to increase the point-cloud object detection accuracy while reducing the cost. Thirdly, we utilize self-supervised training strategies to use unlabeled data efficiently and create a set of geometric pretext tasks to pre-train the neural network. Finally, we study specific challenging real-world scenarios and implement context-specific solutions to problems raised by the presence of work zones.