Object detection/tracking toward underwater photographs by remotely operated vehicles (ROVs)

Abstract

Owing to the shortcoming of image blur, scale diversity, and complex backgrounds, underwater target detection and tracking is a tough challenge in real-world ROVs applications. In this paper, an underwater target detection method based on the class-weighted YOLO network is proposed. The key technique is to construct the class-weighted loss function based on the deep network YOLO to balance the sample difficulty and achieve better results. In addition, an adaptive dimensional clustering algorithm of target box is introduced to further improve the detection performance. Firstly, according to the imaging characteristics of the small underwater target in the side-scan sonar image, the sonar image is filtered and segmented by clustering to significantly reduce noises in each image. Afterward, the speckle detection is deployed to extract the suspected target area in the side-scan sonar image. Finally, the sonar image is segmented according to the automatic threshold segmentation and receive the binary image of the target region. The target scale is estimated by the second moment and the false target is eliminated. Correspondingly, the accurate detection of small underwater target can be achieved. The moving object pixel is detected by background difference and three-frame difference respectively. On this basis, we combine the background difference and the three-frame difference. The background difference characterizes the objects capturing the incomplete message detected by the three-frame difference. Finally, morphological operation is leveraged to eliminate the noise caused by non-static background objects. Experiments on multiple ROVs have validated that, the proposed algorithm can improve the average accuracy and the recall rate remarkably in the dense target detection task that contains 20 targets averagely.