General-Purpose Model-Free Object Tracking with 3D LiDAR

This thesis presents two different methods developed by the author in the area of 3D tracking of multiple objects using 3D LiDAR data. The first method attempts to improve the overall performance of a reference tracking system by using distance-dependent motion modelling. Two of the most common motion models, Kalman Filter and Constant Velocity, are combined using a weight function that depends on the distance between the tracked object and the sensor. As the results show, this technique proved to increase the complexity of the tracking algorithm and its processing load without providing any mayor improvements. On the other hand, the second method was developed with the aim to improve the behavior of 3D multi-object trackers against occlusions and point cloud sparsity. This was achieved by simplifying the life management of the tracklets, consisting of removing both tracklet scoring from the tracklet life or death evaluation and the minimum hits required to generate a track, as well as increasing the number of frames in which a tracklet can survive without being associated with a detection. An important improvement was obtained regarding the number of identity switches for both vehicles and pedestrians. In the pedestrian case, the lowest value seen, at least in the field of 3D model-free multi-object tracking, was obtained, to the best of the author’s knowledge