Transformer Based Advanced Multi-Target Tracking with Guiding Mechanism

In practice, the nonlinearity of the observation model increases the complexity of the Multi-target Tracking (MTT) problem. Compared to model-based methods, deep learning (DL) is able to learn complex relationships in the data through training, and thus is gradually gaining attention in the field of MTT. Different training data can have a significant impact on network models, but there are limited studies that consider the impact of feature differences in nonlinear measurements on network models. To realize an advanced DL-based MTT method for nonlinear systems, we propose a range guiding mechanism that models discrepancies in nonlinear measurement data using target range information, thereby enhancing the robustness and consistency of the features. Our innovative Range Guiding Tracking with Transformer (RGTT) method introduces a range guiding module between the Transformer encoder and decoder to enhance features with range information while regressing the predicted state of the current frame. By continuously analyzing the nonlinear measurements fed into the RGTT, the continuous predicted state of targets in the Cartesian coordinate system can be obtained. Experimental results show that the guiding mechanism enables RGTT to deeply grasp measurement features, resulting in excellent tracking performance and generalization ability.