Towards Data-driven Multi-target Tracking for Autonomous Driving

We investigate the potential of recurrentneural networks (RNNs) to improve traditional on-line multi-target tracking of traffic participants froman ego-vehicle perspective. To this end, we builda modular tracking framework, based on interact-ing multiple models (IMM) and unscented Kalmanfilters (UKF). Following the tracking-by-detectionparadigm, we leverage geometric target propertiesprovided by publicly available 3D object detectors.We then train and integrate two RNNs: A state pre-diction network replaces hand-crafted motion mod-els in our filters and a data association network findsdetection-to-track assignment probabilities. In ourextensive evaluation on the publicly available KITTIdataset we show that our trained models achievecompetitive results and are significantly more robustin the case of unreliable object detections.