A multi-target tracking algorithm suitable for road scenes

Traditionally, multiple target tracking and target detection are carried out using separate models, which leads to non-sharing of features and inability to track in real time. Now with the rapid development of target detection and re-identification (re-ID) technology, this paper proposes a conceptually simple and effective joint detection and tracking model. Its detection module and tracking module are isomorphic networks, which are based on anchor- The CenterNet network of the free mechanism is used as the backbone network, and all the convolutional layers in the up sampling are replaced with deformable convolutions, so as to adapt to the geometric deformation of different tracking targets. Second, the tracking module calculates the re-ID feature for each pixel, outputs the calculated similarity between the detection frame and the trajectory frame, uses the Hungarian algorithm and the correlation metric of the measurable bounding box overlap, and uses the Kalman algorithm to perform data association frame by frame. The trajectory filling strategy is used to balance the number of missed and false positives. Finally, experiments are carried out on the open target tracking data set, which proves the superiority and effectiveness of the proposed model, and realizes the tracking of pedestrians, vehicles, bicycles and tricycles in four target categories.