Object tracking for vehicle

A method comprising:

• associating detected objects from sensor data with tracked objects in a tracking list stored in a memory by executing an association model, the association model including association weights and outputting probabilities that the detected objects are the same as the tracked objects, an association-model gradient of the association model being computable, the association-model gradient being of the probabilities with respect to the association weights;
• determining kinematic states of the tracked objects by executing a Bayesian filter, the Bayesian filter including Bayesian parameters, a Bayesian-filter gradient of the Bayesian filter being computable, the Bayesian-filter gradient being of the kinematic states with respect to the Bayesian parameters;
• and
• updating which of the tracked objects are stored in the tracking list by executing a track-management model, a track-management gradient of the track-management model being computable;
• wherein the association model, the Bayesian filter, and the track-management model are jointly trained.

A computer comprising a processor and a memory, the memory storing instructions executable by the processor to:

• associate detected objects from sensor data with tracked objects in a tracking list stored in the memory by executing an association model, the association model including association weights and outputting probabilities that the detected objects are the same as the tracked objects, an association-model gradient of the association model being computable, the association-model gradient being of the probabilities with respect to the association weights;
• determine kinematic states of the tracked objects by executing a Bayesian filter, the Bayesian filter including Bayesian parameters, a Bayesian-filter gradient of the Bayesian filter being computable, the Bayesian-filter gradient being of the kinematic states with respect to the Bayesian parameters;
• and
• update which of the tracked objects are stored in the tracking list by executing a track-management model, a track-management gradient of the track-management model being computable;
• wherein the track-management model is a Markov decision process;
• the Markov decision process includes a reward function defining utilities of transitioning between states and a policy function defining actions based on the states;
• and
• the track-management gradient is of the reward function with respect to the policy function.