Learning to Multi-Target Tracking in Dense Clutter Environment with JPDA-Recurrent Neural Networks

Abstract

Multi-target tracking in dense clutter environment is an open problem and still involves many challenges. Different from the classical tracking methods involving complex models and accurate prior knowledge, deep learning methods have been researched for this problem in recent years. In this paper, a novel JPDA-recurrent neural networks (RNNs) based tracking approach is proposed. Three stages which can address data association, state prediction and track management are included in our approach. For the prediction stage, a model for modeling temporal sequence based on long-short term memory (LSTM) is employed to learn the targets motion parameters from radar noisy measurements. Moreover, an RNN-based track existence probability model is proposed to assess the track quality and to automatically initialize, maintain and terminate the tracks. This approach is demonstrated over a simulation scenario, with promising results.