Stern target tracking algorithm based on data fusion

This paper proposes a stern target detection and tracking algorithm based on multi-sensor decision-level fusion. Aiming at the problems of excessive noise and difficult target segmentation in Lidar point cloud data, bilateral filtering and voxel filtering were adopted for preprocessing, and an adaptive parameter DBSCAN clustering method was designed to achieve accurate point cloud segmentation and three-dimensional bounding box fitting of the stern target. Combining the visual and point cloud data fusion strategy based on projection geometry and matching the detection results through the IoU threshold of 0.9, the accuracy of target recognition has been significantly improved. Finally, multi-target tracking is achieved based on the improved DeepSORT algorithm. Experiments prove that on the self-built ship dataset, this method maintains a similar tracking accuracy to the original DeepSORT (MOTA is $\mathbf{7 3. 1 \%}$, MOTP is $\mathbf{6 3. 6 \%}$, IDSW is 36) while increasing the running speed to 63 FPS. It is significantly higher than the 45 FPS of DeepSORT, demonstrating better real-time performance and practicability.