Classification Assisted Tracking for Autonomous Driving Domain

Augmenting the reliability of extended object tracking is a major issue to enable autonomous driving. Therefore, classification assisted tracking is introduced. A new multiple extended object tracking filter using radar measurements from a non synchronized radar network with classification results from a deep neural network. Formulas are derived to calculate association scores using classification results and radar measurements of extended objects with which global nearest neighbor measurement to track association is performed. In this paper an random matrix interacting multiple motion model is proposed for single state estimation on a non synchronized sensor network. A dual hidden Markov model is introduced to determine the object's classification and motion state which are needed to calculate data association scores. Classification assisted tracking is tested on a real world scenario and evaluated with the multiple object tracking accuracy (MOTA).