An evaluation of dynamic object tracking with 3D LIDAR

Abstract

This paper evaluates experimentally approaches to tracking dynamic objects using 3D LIDAR data such as that obtained from a Velodyne HDL-64. Whilst many algorithms exist for tracking objects in 2D laser data, only recently have algorithms which model the 3D appearance of tracked objects been proposed. The benets of these algorithms with respect to model free centroid based methods have not been fully examined, and therefore the key contribution of this paper is to analyse the behaviour of dierent implementations of these two broad approaches. Through a quantied comparison of simple tracking methods for 3D LIDAR data validated against manually labelled tracks it was found that centroid tracking outperformed 3D appearance model based techniques. In the experiments, which focused upon pedestrian tracking in particular, centroid tracking achieved up to 95% correct data associations compared to 60% for the appearance model based methods.