TrackLidFormer: A Transformer-Based Approach for Occluded Object Tracking

Abstract

Through the growing complexity and functionality of automated driving functions, also the requirements for testing and evaluation grow continuously. This leads to an increasing interest in virtual test drives for evaluation purposes. Detecting and tracking dynamic objects, like cars and pedestrians, in LiDAR sensor data for Autonomous Vehicles is vital. Especially for the derivation of long-term scenarios and driver behavior models for use in simulation, objects need to be tracked consistently from the initial to the final frame. In this paper, we propose a multi-object tracking model employing an encoder-decoder transformer architecture, utilizing detected object features from past frames as the encoder input and current frame object features as the decoder input to generate object trajectories. This approach stands in contrast to conventional object detectors and trackers, which often limit the association to a few past frame objects, failing to associate long-term occluded objects. In comparison, our proposed method achieves successful tracking, even in critical scenarios. Evaluations on the nuScenes dataset demonstrate competitive performance in LiDAR-only tracking tasks, contributing to the development of more realistic simulations for the evaluation of autonomous driving functions.