Real-time Detection and Tracking Network with Feature Sharing

Multiple object tracking (MOT) systems can benefit many applications, such as autonomous driving, action recognition, and surveillance. State-of-the-art methods detect objects in an image and then use a representation model to connect these objects with existing trajectories. However, the combination of these two components to reduce computation has received minimal attention. In this study, we propose a single-shot network for simultaneously detecting objects and extracting tracking features to achieve a real-time MOT system. We also present a detection-tracking coupled method that uses temporal information to improve the accuracy of object detection and make trajectories complete. Experimentation on the KITTI driving dataset indicates that our scheme achieves an accurate and fast MOT system. In particular, the lightweight network reaches a running speed of 100 FPS.