Ship multiobject tracking method based on DeepSORT

To address the occlusion problem in ship target tracking, this paper proposes an improved Deep Simple Online and Realtime Tracking (DeepSORT) algorithm by associating it with Automatic Identification System (AIS) information. First, the historical trajectory of all targets contained in AIS data and detections is extracted. Then, the cost matrix is constructed by calculating the trajectory similarity, and the Hungarian algorithm is used to complete the optimal allocation. When the ship is blocked, the corresponding AIS information is extracted to construct state vector, and the state vector is weighted with the tracker prediction information to update the Kalman filter. Aiming at the failure of appearance feature matching during partial occlusion, Scale-Invariant Feature Transform (SIFT) feature point matching method is added in this paper to optimize the matching strategy. In order to verify the effectiveness of the proposed method, experiments were conducted on 24 videos of passing ships in the Wuhan section of the Yangtze River, which contained corresponding AIS information. The results showed that the proposed method performed well and exceeded the original method in multiple indicators, and ID switch decreased significantly. This indicates that the method proposed in this paper has a higher retention ability for ship identification, and is expected to help maritime supervision.