The use of high-resolution satellite video data for real-time detection and tracking of large-scale maritime targets has important applications in national defense, military, public safety, and other fields. However, existing research on object tracking mainly focuses on the field of natural images and mainly addresses single object tracking problems. A multi task learning model based on point representation for object detection and identity re-identification is proposed to address the problems of complex sea surface backgrounds, diverse ship sizes, and small pixel occupancy of some targets. The model uses a multi-level feature aggregation network in the CenterNet network to extract multi-scale features, and achieves object detection through the recognition and localization of the target center point, avoiding the computational overhead of anchor and non-maximum suppression. In addition, the use of target identity classification and multi task learning loss adaptive balancing methods has solved the problem of feature learning target conflicts in ship target detection and identity re-identification tasks. By building a target tracking dataset and testing it on 24 videos, the target detection recall rate, target detection accuracy rate, multi-target tracking recall rate, and multi-target tracking accuracy rate of the multi-target tracking system reached 88.3%, 95.9%, 90.2%, and 98.0%, respectively, achieving good target tracking.