Abstract: ‍ ‍Clutter and dense targets often cause association errors as tracking environments become increasingly complex. The optimal matching process grows exponentially with the number of targets， highlighting the complexity and unreliability of traditional algorithms for multi-target track association. This study proposes a deep association network based on probability hypothesis density （PHD） map features and sequence features to address optimal matching issues caused by varying target numbers and clutter interference. First， the PHD map of targets was fused with the PHD map of measurements to explore spatial association features and structural information. Then， point features were extended to image features， increasing the basis for association discrimination. The study developed a new measurement method based on deep learning for point or sequence features， reducing the dependency on distance and thereby mitigating misjudgments caused by dense clutter. Finally， a deep matching network based on a multi-head attention mechanism was proposed to avoid combinatorial explosion as the number of targets and measurements increases. This network used an eight-head attention mechanism to focus on neighborhood distribution， improving the accuracy and reliability of the matching process while reducing computational time. Ablation experiments demonstrated that image features， sequence features， and the deep matching network enhance the feature extraction capabilities for association information and the global optimal matching process. Simulation experiments showed that this deep association network improves the accuracy and stability of association in both the affinity matrix and the minimum assignment process. Additionally， this method illustrates the feasibility of transferring existing image tracking association algorithms to radar tracking， providing new insights for developing radar tracking data association.