Research on Distribution Strategy of Region of Interest in Panoramic Video Based on Improved Deepsort

The difficulty of data transmission of panoramic video has always limited the wide application of panoramic video. In order to solve this problem, a panoramic distribution algorithm based on the region of interest is proposed. Since the user's focus is often on the person appearing in the video, we consider the area around the person as a potential area of interest. The acquisition of the region of interest is based on a deep learning method. YOLOv4 is used to perform target detection on the panoramic video, and the detected pedestrian target is input into the multi-target tracking module based on the Deepsort algorithm. We use the center of the bounding box output by the Deepsort algorithm as the center of the region of interest and distribute the region of interest in the panoramic video. This greatly reduces the amount of data transmitted by panoramic video, and basically does not affect the user's viewing experience. In order to reduce the switching of target id, this paper improves the appearance extraction model in Deepsort, and proposes an appearance extraction network that focuses on both local and global information, so that the appearance extraction model can extract more representative appearance features of pedestrians. The validity of the model is verified on multiple pedestrian re-identification data sets.