3D Multi-Object Tracking With Adaptive Cubature Kalman Filter for Autonomous Driving

Abstract

A crucial and challenging issue in autonomous driving is dynamic road environment detection and 3D multi object tracking. In this article, we propose a novel framework for online 3D multi-object tracking to eliminate the influence of inherent uncertainty and unknown biases in point cloud. A constant turn rate and velocity (CTRV) motion model is employed to estimate the future motion state, which are smoothed by a cubature Kalman filter (CKF) algorithm. A new affinity model is introduced to evaluate the similarity between trajectories and candidate detections for accurate and reliable data association which can be formulated as a bipartite matching problem. An adaptive cubature Kalman filter (ACKF) is given to remove the influence of unknown bias and to robustly update the tracked state. Accuracy and speed of the proposed tracking method are evaluated on the KITTI 3D multi-object tracking dataset, showing superior performance than the baselines.