Research on Crowd Tracking Methods

Crowd tracking is an important research topic in the field of computer vision. In crowd scenes, tracking pedestrians often results in missing individuals, i.e. individuals without trajectory information, and incorrectly identifying pedestrian individuals, i.e. identifying non pedestrian targets as pedestrian individuals for tracking. Therefore, it limits the tracking accuracy of existing methods. In the past, most methods only considered pedestrian tracking research in sparse scenes, so they often overlooked the importance of these two issues on crowd tracking. To solve these two problems, this paper proposes a density map based crowd tracking method. This method includes three task branches, namely pedestrian individual detection branch, density map estimation branch, and re identification branch. The pedestrian individual detection branch is responsible for predicting the position information of pedestrian individuals. The density map estimation branch can obtain the number of pedestrian individuals in the image area, which is used to jointly construct the mean square error loss with the pedestrian individual detection branch. In crowded scenes, this method can generate a constraint on the number of pedestrian individuals in the detection branch by counting the population density map, which can help the pedestrian individual detection branch find missed individuals. Similarly, the heatmap of the pedestrian individual detection branch can be used to enhance the localization and counting ability of the population density map, forming a bidirectional constraint. The re identification task branch is responsible for extracting the apparent features of individual pedestrians, which are combined with their location information to establish continuous inter frame pedestrian connections. At the same time, in response to the problem of incorrect recognition of pedestrian individuals during the tracking process, which leads to incorrect association of the same individual between consecutive frames and affects overall tracking performance, this paper improves the DLA feature extraction network by optimizing the network's upsampling and downsampling methods, improving the model's detection accuracy for pedestrian individuals and enhancing tracking performance. Through experimental verification, the MOTA and IDF1 on the MOT20 dataset were 78.8% and 82.6%, respectively. The more advanced method ByteTrack improved the MOTA and IDF1 by 1% and 7.4%, respectively.