Multiple Vehicle Tracking Based on Labeled Multiple Bernoulli Filter Using Pre-Clustered Laser Range Finder Data

Abstract

Multiple vehicle tracking (MVT) system is a prerequisite to path planning and decision making of self-driving cars as it can provide positions of surrounding vehicles. Most of the available approaches belonging to the so called tracking-by-detection approach inevitably bring detection errors into the tracking result. In this study, we proposed a laser range finder (LRF) based track-before-detect MVT algorithm without detection procedure. Moreover, different from the state of the art in track-before-detect approaches using raw data, we applied a pre-clustering procedure to segment the raw data into disjoint clusters to reduce computation demand. Specifically, a clustering algorithm named iterative nearest point search (INPS) which can even handle the partial occlusion situations that are challenging for traditional clustering algorithms was designed for the pre-clustering procedure. Furthermore, a detailed cluster-to-target measurement model was proposed to describe the difference between cluster and hypothesis vehicle. Finally, we integrated the measurement model into the labeled multi-Bernoulli filter with particle implementation. Simulations and experiments show that the proposed MVT algorithm provides more accurate estimates of vehicle number and position in comparison with conventional methods.