Multi-Target Trajectory Prediction Algorithm Based on Spatiotemporal Feature Fusion and Attention Mechanism

In radar multi-target tracking， the robustness of the motion model directly affects overall tracking performance. As the tracking environment becomes increasingly complex， weakly constrained， non-cooperative target tracking is prone to challenges such as target occlusion， measurement loss， and unpredictable target motion paths. This paper proposes a region proposal net-attention-convLSTM-transformer （RPN-AT-ConvLSTMT） spatiotemporal feature extraction and prediction network for constructing motion models of weakly constrained， non-cooperative targets. First， one-dimensional radar measurements are extended into a two-dimensional Probability Hypothesis Density （PHD） map using covariance and mean. This enables a probabilistic representation of the regions where targets may exist. To address the challenge of large search areas in multi-target tracking， preprocessing operations such as multi-scale feature transformation and key area annotation are applied to the image. These steps help identify local regions or search areas for tracking and reduce the network’s focus on irrelevant regions. Furthermore， an improved multi-head attention mechanism based on the Region Proposal Net （RPN） is designed. This mechanism dynamically reconstructs the attention weight matrix using Gaussian distribution characteristics， target motion priors， and interaction relationships. It ensures a dynamic balance between global and local regions， enhancing the attention mechanism’s focus on target-relevant areas. The parallel architecture of the Transformer encoder enables deep fusion and efficient transmission of spatiotemporal features. Finally， a BP （Back Propagation） neural network is employed to extract the Gaussian distribution’s mean and covariance from the predicted PHD map， completing the target state estimation. Ablation experiments verify the effectiveness of the proposed attention mechanism and spatiotemporal fusion module. Simulation results in multi-target tracking scenarios demonstrate that the proposed algorithm significantly improves tracking accuracy compared to the ConvLSTM method， particularly in cases of occlusion and measurement loss， offering a robust solution for weakly constrained， non-cooperative target tracking.