关键词: 目标跟踪, 交互式多模型, 可学习膨胀因果卷积, 全局-局部门控注意力, 深度学习

Abstract: In target tracking, the dynamic characteristics of highly maneuvering targets pose significant challenges to the robustness and accuracy of tracking algorithms. Traditional Interactive Multiple Model algorithms employ fixed model parameters and transition probability matrices, which often lead to response delays and large transient errors during motion model switching. In addition, their limited capability in exploiting temporal features of input sequences restricts tracking performance in complex scenarios. To address these issues, this paper proposes a DCGLA-IMM algorithm based on the BiLSTM-IMM framework, in which deep temporal modeling and attention mechanisms are introduced to enhance feature extraction and dynamic adaptability. Specifically, causal convolution with learnable dilation rates is utilized to extract multi-scale local temporal features, while a two-layer BiLSTM with residual connections is employed to strengthen long-range temporal modeling. Furthermore, a global-local gated hybrid attention mechanism is incorporated to achieve adaptive fusion of local details and global trends. Finally, a GRU integrates sequential information, and multiple prediction heads output the probabilities of different motion models, followed by probability correction and state fusion. Simulation results demonstrate that, under two challenging scenarios with time-varying parameters of fixed-structure motion models and multi-model switching, the proposed algorithm reduces the RMSE by 10.07% and 9.14%, respectively, compared with the BiLSTM-IMM. It also achieves superior performance in terms of model probability response delay and probability quality metrics, thereby validating its improved accuracy and robustness for tracking highly maneuvering targets.

Key words: maneuvering target tracking, interacting multiple model, learnable dilated causal convolution, global-local gated attention, deep learning