Transformer-based two-source motion model for multi-object tracking

摘要

Recently, benefit from the development of detection models, the multi-object tracking method based on tracking-by-detection has greatly improved performance. However, most methods still utilize traditional motion models for position prediction, such as the constant velocity model and Kalman filter. Only a few methods adopt deep network-based methods for prediction. Still, these methods only exploit the simplest RNN(Recurrent Neural Network) to predict the position, and the position offset caused by the camera movement is not considered. Therefore, inspired by the outstanding performance of Transformer in temporal tasks, this paper proposes a Transformer-based motion model for multi-object tracking. By taking the historical position difference of the target and the offset vector between consecutive frames as input, the model considers the motion of the target itself and the camera at the same time, which improves the prediction accuracy of the motion model used in the multi-target tracking method, thereby improving tracking performance. Through comparative experiments and tracking results on MOTchallenge benchmarks, the effectiveness of the proposed method is proved.