Using Feed-Forward Neural Networks for Data Association on Multi-Object Tracking Tasks

Abstract

This article presents an approach for data association in single camera, multi-object tracking scenarios using feed-forward neural networks (FFNN). The challenges of data association are object occlusions and changing features which are used to describe objects during the process. The presented algorithm within this article can be applied to any kind of object which has to be tracked, e.g. persons and vehicles. This approach arises within a project to detect critical behavior of persons. Besides, person tracking is one of the most challenging scenarios. People have different velocities and often change the moving direction. In addition, a variety of occlusions are caused by the movement as a group. Also in most surveillance scenarios the illumination conditions are not optimal. The usage of a feed-forward neural network is a mostly new approach in this research field. The advantage is the lightweight computational complexity and the fixed termination time in contrast to recursive neural networks like Hopfield networks which are used for plot association during radar tracking. FFNN is a non-probabilistic approach in contrast to common algorithms within this filed. They deliver decisions not probability values. The handling of the FFNN output will be presented in this article. During the evaluation we will show that the developed approach is capable to handle completely different scenarios like tracking people moving mostly straight forward but also complex scenarios like a soccer game.