Optimization Method of Dual-Decoder Transformer Model in Multi-Target Tracking

In view of the multi-target tracking algorithm in complex scenes, the multi-target pedestrian tracking is affected by factors such as dense pedestrians and light changes, which makes the multi-target pedestrian tracking have problems such as false detection and missed detection, identity change and inaccurate association. This paper introduces a Transformer-based dual-decoder network for multi-pedestrian tracking. First, the Mish activation function is incorporated into the feature extraction ResNet network to mitigate the gradient vanishing issue, thereby enhancing the model's performance and generalization. In the target detection module, a selective query memory mechanism is employed to improve the decoder's ability to capture essential features more effectively, thereby alleviating the problem of false detection and missed detection of targets. In the data association part, the enhanced secondary association method is applied to address inaccurate associations caused by brief occlusions in dense pedestrian crowds. Experimental results on the MOT17 dataset demonstrate that, compared to the TransTrack model, the proposed network improves the MOTA by 4.6%, the IDF1 identity score by 4.9%, and AssA by 5.1%.