Persistent 3D Multi-Object Detection and Tracking Using Sensor Fusion for Autonomous Vehicles

Abstract

Persistent multi-object tracking (MOT) allows autonomous vehicles to navigate safely in dynamic environments. One of the common challenges in MOT is object occlusion, when an object becomes undetectable for subsequent frames. Even though the current MOT methods temporarily maintain objects’ trajectories in memory to overcome the issue, they retain short-term memory to avoid unnecessary associations that aggravate the computational speed and eventually slow down the overall solution. As a result, they lose track of some objects during occlusions, particularly long ones. This research investigates this issue by involving subsequent experiments on MOT techniques and locating defects that raise the issue. In addition, we propose an MOT framework that recovers objects after occlusions. Initially, we study the pros and cons of object tracking by features that reflect their appearance and evaluate the performance under occlusion scenarios. Then, we implement a similar study for object tracking by trajectory estimation using a naive motion model and Kalman filters. Given the defects observed in the study, we propose a light MOT method that utilizes data from a camera and LiDAR sensors and performs association and fusion in an algebraic form, which enhances the computational speed and permits long-term memory. Our method shows outstanding performance over recent learning and non-learning benchmarks with about 3% and 4% margin in MOTA, respectively. We conduct extensive experiments that simulate occlusion phenomena by employing detectors with low/moderate/high distortions, where our method achieves superior performance over current state-of-art methods. Our framework processes about 7,763 frames in 1.48 seconds, which is seven times faster than recent benchmarks.

Date of Award	Dec 2022
Original language	American English
Supervisor	Majid Khonji (Supervisor)