An Analysis and Benchmarking in Autoware.AI and OpenPCDet LiDAR-based 3D Object Detection Models

Abstract

Light Detection And Ranging (LiDAR) sensors are widely used in applications related to autonomous driving. The ability to scan and visualize the 3D surroundings of the vehicle as a point cloud is a particular strength of this sensor. Various different object detection models have been proposed to provide bounding box predictions given a point cloud. This thesis looks at two popular, open-source frameworks which provide solutions to this problem, Autoware.AI and OpenPCDet. The Autoware.AI framework provides models which use hand-crafted, non-neural network based methods to solve LiDAR-based object detection, while the OpenPCDet framework provides models based on neural networks. In this thesis, these models are compared with each other on a custom labeled dataset. As expected, the results of this comparison show that the non-neural network based Autoware.AI models perform significantly worse than the neural network based OpenPCDet models. Additionally, it is shown that amongst the OpenPCDet models, PV-RCNN performs better for detecting vehicles, SECOND and PV-RCNN perform better for detecting pedestrians, and SECOND and Part-A^2 Free perform better for detecting cyclists.