Underwater Target Tracking Technology Based on YOLO v4 and Deepsort

Underwater target tracking is a technology that tracks the motion of objects in underwater environment and provides accurate position information. With the progress of online real-time tracking technology, the accuracy of underwater target tracking has also been improved, and the time consumed in the calculation process has also been reduced. However, the poor underwater imaging conditions lead to the temporary disappearance of objects and serious position occlusion. At the same time, underwater imaging will make the object change in shape and size to a certain extent. Poor imaging conditions will make the result of target tracking unsatisfactory. Therefore, this paper uses the combination of YOLO v4 and SORT WITH DEEP ASSOCIATION METRIC (Deepsort) to complete the underwater target tracking task. YOLO v4 algorithm uses CSP Marknet 53 as the backbone network, which improves the parameter fitting ability of CNN and maintains the lightweight structure of CNN. The Spatial Pyramid Pooling (SPP) structure adopted by YOLO v4 can improve the receptive field of CNN and improve the recognition accuracy of small objects. Deepsort uses Mahalanobis distance to measure the distance between object detection Bbox and tracking Bbox, which can more accurately measure the relevance of object position information. At the same time, the feature corresponding to the Bbox calculated by CNN is used to measure the correlation of objects, reduce the occurrence of re-recognition phenomenon, and achieve better tracking effect. Finally, the test results in the underwater data set are MOTA 61.1 and Runtime 60Hz.