A Novel Vehicle Tracking Method for Cross-Area Sensor Fusion with Reinforcement Learning Based GMM

Radars, LiDARs and cameras have been widely adopted in autonomous driving applications due to their complementary capabilities of environment perception. However, one problem lies in how to effectively improve the cross-area tracking accuracy with massive data from multiple sensors. This paper proposes a novel tracking solution that is composed of a reinforcement-learning-based Gaussian mixture model (GMM), submodel center realignment, and data-driven trajectory association. Specifically, developed with a Q-learning-based cluster number, an improved GMM-EM algorithm is firstly investigated to cluster the dense short-range radar data points. Subsequently, an innovative kinetic-energy-aware approach is presented to realign the Q-learning GMM cluster centers for position error mitigation. In addition to Q-learning GMM clustering, a weight-scheduled method is presented to associate the data from a long-range radar and cameras for cross-area object extraction and trajectory fusion. Eighteen experiments for training and one experiment for verification were conducted on a fully-instrumented autonomous vehicle. Experimental results demonstrate that a better tracking performance in crossing detection areas can be achieved by the proposed method.