一种基于YOLO-V3算法的水下目标识别跟踪方法

In order to assist the underwater platform to complete the autonomous shooting task, a target recognition model based on the YOLO-V3 algorithm is proposed. The network structure is optimized by the means of down-sampling recombination, multi-stage fusion, optimization of clustering candidate box, and redefinition of loss functions, etc, which improves the accuracy of target recognition and the calculation speed of the algorithm. The feature description method with rotational invariance is applied to track the motion of multi-degree of freedom objects in water, and the tracking state is corrected by evaluation results. Experiments show that the method can identify and track the target autonomously and has adaptive ability. For the image with input pixel of 416*416, the processing speed reaches more than 15 frames per second and the mean Average Precision (mAP) reaches 75.1 when the confidence degree is 0.5, which meets the real-time and accuracy requirements.