Multi-object tracking using feed-forward neural networks

In this article we present an approach for robust multi-object tracking. Typically the task of object tracking can be divided into two subtasks: object detection and object labeling. The main focus of the work presented here is on an approach for consistently labeling objects across a series of video frames using neural networks. Due to the specialization of object detection algortihms it is necessary to divide detection and labeling to enhance their individual skills. In the evaluation we show that the developed labeling is robust against occlusions and can handle low object detection rates.