Improved Joint Probabilistic Data Association for multiple target tracking in clutter based on Imitation learning

The input of a multi-target tracking system is the data detected by the sensor, which includes observations of the correct target and clutter fromthe tracking space. In order to find the appropriate observation for each track, it is necessary to associate the sensor data. The presence of multiple targets along with the increase of noisy observations increases the order of the calculations and reduces the tracking accuracy. The Joint Probabilistic Data Association Algorithm (JPDA) is an effective algorithm for solving the problem of simultaneously associating multiple observations to multiple targets by generating all possible hypotheses. In general, hypothesis means assigning a maximum of one observation to each track in such a way that none of the observations inside the hypothesis are assigned to track more than once. Despite the optimal performance of the JPDA algorithm, the increase in noisy observations and the intersectionof the tracks generatestoo manyassignment hypotheses, decrease in accuracy, and slow down and even stop the tracking process. The algorithm proposed in this article by using the imitation learning algorithm with the real-time training process and removing the hypothesis generation step when there is a possibility of explosion of dimensions in the JPDA algorithm, alongmaintaining the approximate accuracy of estimating the states of each target, increase the speed of the assignment in the presence of noise and clutter observations. In this algorithm, the Imitation learning model has been mediated in order to extract the required information from the sensor observations and has eliminated the need for hypothesis generation in general. The simulation results show that the proposed algorithm, while maintaining the tracking accuracy in comparison to JPDA, has a lower calculation than the JPDA algorithm. In the situation in whichthe JPDA faces the problem of computational dimension explosion due to the overlap of target observations or the number of clutters, the proposed approachusing the learned model shows a suitable performance and prevent the dimension explosion.