A novel multi-target track initiation method based on convolution neural network

Abstract

This paper addresses the problem of track initiation for multi-target in different motion forms and in complicated clutter background. The method proposed combines the traditional logic-based method and convolution neural network. The logic-based method is used mainly to generate a set of track proposals, which is computed by the convolution neural network to extract features in data domain. In this paper, softmax at the end of the convolution neural network is substituted by a one-dimensional two-class classifier for the output layer of the convolution neural network is designed to output a one-dimensional value. There are two key insights in this method: (1) the classification problem has been transformed into target tracking problem on the condition that the set of track proposals is found. (2) the convolution neural network is firstly used in data domain to mine and augment high-level features that make classification more easily. The simulation experiments have shown that this method performs much better than modified Hough transform which is used to initialize tracks traditionally, especially when the targets are maneuver. In the experiments based on real data, this method is proved to be adaptive enough to initialize tracks whose data comes from different radars.