A Deep Learning Approach to the Detection and Tracking of Moving Objects in 2D Point Clouds

The detection and tracking of moving objects (DATMO) are crucial tasks that any autonomous vehicle must perform. Autonomous vehicles must detect and track all obstacles to ensure safety within the environment while also completing their tasks efficiently. In autonomous driving research, LiDAR is becoming increasingly popular due to its high resolution and accuracy. There are many state-of-the-art DATMO methods using LiDAR, however, most methods are designed for 3D LiDAR sensors. Methods that work for 2D LiDAR sensors are not as robust as their 3D counterparts or require too many computational resources to run efficiently on less powerful robots. This research presents two robust solutions to the DATMO problem based on deep learning techniques that can scale to meet a variety of hardware constraints. The first solution, detect while track (DWT), combines a convolutional neural network (CNN) with a multiple hypothesis tracking (MHT) approach and Kalman filter. The second solution, pixel predictions for future-oriented bounding boxes (PIXFOR), combines a CNN with a recurrent network architecture to solve both detection and tracking problems in a single forward pass. Both methods are experimentally validated on an unmanned ground vehicle (UGV) operating on an intersection scenario and a highway scenario using 2D point clouds collected from simulation and hardware environments. The run-time performance of both methods is also validated different hardware platforms to show that the methods can scale to meet different hardware constraints. When compared to state-of-the-art DATMO methods, the newly proposed methods outperform in the object detection and tracking tasks, while operating at a faster run time on equivalent hardware.