Online Multi-object Tracking with Siamese Network and Optical Flow

Abstract

Simple Online and Realtime Tracking with a Deep Association Metric (DeepSORT) is a method of multi-object tracking combined appearance features with motion state of objects estimated by Kalman Filter which has a promising performance. However, maintaining the identity of targets becomes formidable when the objects have a similar appearance and complex patterns of the movement. To address these issues, a novel Online Multi-object Tracking with Siamese Network and Optical Flow is proposed. We utilize the Siamese network structure to obtain our appearance feature extractor. Furthermore, optical flow is introduced into the scheme to promote the accuracy of motion prediction from the Kalman filter. Our approach combines appearance and motion features in a tracking framework. The experimental results evaluated on the public MOT dataset illustrate that our method has the better performance in comparison with the DeepSORT algorithm.