Improving The Tracking Persistence of Multi-object Tracking using Scene Classification

In this paper, we track swimmers in an image dataset by using the multi-object tracking (MOT) network called FairMOT. There are problems when tracking swimming datasets through this network, and to address those problems we added a scene detection network to classify input images and to adjust optimized weights and hyper-parameters according to the classes. In addition to the existing IoU formula, we improved performance by assigning optimized weights to each class based on the aspect ratio, considering the center point distance and the aspect ratio between bounding boxes. Additionally, we found optimized values for the hyper-parameters, position, and velocity of the Kalman filter for each class. Utilizing the classifications by the scene detection network, we adjusted the optimized hyper-parameters based on the scene, resulting in a maximum improvement of 5.2% in multi-object tracking accuracy scores.