Multi-object tracking using context-sensitive enhancement via feature fusion

Multi-object tracking (MOT) is one of the most challenging tasks in the field of computer vision. Most MOT methods generally face the problem of not being able to handle pedestrian features such as size and appearance well, which can easily lead to the problem of missed detection and occlusion. Considering this, an end-to-end multi-target tracking network with feature fusion and feature enhancement is proposed. The network framework integrates feature extraction, object detection, and data association. Using two adjacent frames as input chain nodes, based on Inception convolution as the backbone network, which has special pre-training weights that increase the perceptual domain of the network for multiple targets. In addition, the three-times repetitive overlay weighted bidirectional pyramid structure in the feature fusion module, which can focus more on key features and enhance the adaptability to target deformation. In order to solve the phenomenon of crowding in complex scenes, a context-sensitive prediction modules are added, which contain deeper and wider convolution to enhance the key information between targets. After the above processing, three loss function branches are formed, where the classification branch and the identity branch together form the attention multiplied by the regression branch to ensure the accuracy of regression. In MOT16 and MOT17 dataset experiments, our model MOTA metrics reach 67.9 and 67.7, with frame rates up to 30 FPS on a single GPU, with improved visualization results beyond Chain-Tracker.