千兆图像序列的多目标跟踪

Multi-object tracking of giga image sequences is a challenging computer vision task, and the GigaVision Challenges provide image data from giga cameras in a real-world scene for the project. In this paper, a multi-target tracking method for giga image sequences is implemented based on the DeepSORT framework, and the new problems in multi-target tracking of gigab image sequences are solved in multiple strategies. In order to solve the problem of excessive giga image size, an object detection strategy with multi-scale image cropping is adopted. The double-threshold scaling improved trajectory association process is designed. A strategy of using weighted box fusion algorithm to suppress the object detection frame of image block boundary is proposed. The feature matching order in the process of object and trajectory association is improved to solve the problem of low frame rate of giga image sequences. The method ranked 8/28 in the multi-target tracking task of the 2022 GigaVision Challenges.