MM-Tracker: Visual Tracking With A Multi-Task Model Integrating Detection and Differentiating Feature Extraction

Visual tracking is a vitally important task in computer vision, which is widely used in intelligent surveillance and traffic control, etc. Currently, real-time multiple object tracking methods are still not mature in practical applications and still need to be further refined to enhance their performance especially in complex and crowded environments. These existing tracking methods focus on designing the strategies of target association, based on the predictions of object detectors. Since the detection and association phases are independent, many duplicate calculations may occur in the feature extraction procedure, which is time-consuming. Aiming to effectively combine these two phases, an integrated network architecture is proposed in this work to extract the common features for object detection and the differentiating features for target association simultaneously. The multi-task head structure and the specific loss function of this integrated network have been designed to avoid coupling issues in the model training procedure. Furthermore, sample groups of adjacent frames are used to enhance the processing capacity of the personal gradual variance of these tracked targets. Background modeling and a Kalman filter are used to improve tracking accuracy. In the simulation, the proposed tracker achieves better performance than other state-of-the-art trackers on the tested publicly available databases.