Deep Learning-Based Track Prediction and Correction for a Radar Target

Tracking targets using noisy observations is of crucial importance for various civilian and military applications. This involves, among other tasks, ascertaining the state of a target at a next time instant, called as track prediction, and combining the predicted state with a noisy observation to obtain an updated state at the current time instant, called as track correction. We consider track prediction and correction tasks of a radar target tracking application in a highly manoeuvring scenario where the targets exhibit motion in a 3D space under randomly changing constant velocity, acceleration and turn models. The detections from the targets are available at irregular time intervals. We pose the problem of prediction (correction) as a time-series regression problem where several past detections along with their time stamps are used as input, and the next (respectively, current) detection is used as the expected output. We obtain temporal convolutional network (TCN) based models for prediction and correction, which are shown to perform significantly better than the classical pre-fixed interacting multiple model (IMM) algorithm under various different scenarios in terms of the root mean squared error and standard deviation of the error between the true and predicted/corrected values, especially during model transitions and ascend/descend phases with manoeuvres.