A Transformer-Based Unified Framework for Multiple Target Track Association and Fusion under Incomplete Measurements

Multi-sensor multi-target track segment association and fusion play a critical role in maritime target surveillance systems; however, it is difficult to balance feature extraction efficiency, multi-task adaptability and kinematic state estimation accuracy under complex maneuvering scenarios. To tackle this problem, we propose an end-to-end Transformer-based multi-target track association and fusion framework (TMTAFF), which is an offline method. TMTAFF mainly comprises three core components: Track Completion Module (TCM), Track Association Module (TAM) and Track Fusion Module (TFM). The designed TCM is first used to complete the interrupted track segments, and the completed track segments are further fed into the TAM to accurately determine whether any two track segments originate from an identical target. Based on the outputs of the TAM, the TFM subsequently fuses the track segments of each target and outputs the final track fusion results. It is noted that TMTAFF can be simultaneously applied to both interrupted track segment association (ITSA) tasks and multi-source track segment association (MSTSA) tasks due to the adoption of the TCM. Experimental results based on a public dataset show that TMTAFF outperforms 11 state-of-the-art baselines, including traditional approaches and neural network-based models.