水下鱼类目标智能跟踪识别系统的设计与实现

Design and Implementation of Underwater Fish Target Intelligent Tracking and Recognition System

当前通过水下摄像头采集到的水下鱼类目标视频往往存在失真、噪声干扰及光线折射等问题,而传统的视频模式识别跟踪算法在检测复杂、模糊、重叠和紧凑的目标时结果普遍差。基于TensorFlow 2.0和YOLO V4,结合深度学习,设计1个水下视频鱼类目标智能跟踪识别系统。系统对6种鱼类视频的识别准确率(Mean Average Precision,mAP)达到98.50%,跟踪效率(Frames Per Second,FPS)达到46。测试数据表明:系统实现了水下鱼类视频目标的准确识别和实时跟踪,为水产养殖的科学化与数字化提供了智能支持。

At present,the underwater fish target video collected by underwater camera often has problems such as distortion,noise interference and light refraction,while the traditional video pattern recognition and tracking algorithm generally has poor results in detecting complex,fuzzy,overlapping and compact targets.Based on TensorFlow 2.0 and YOLO V4,combined with deep learning,an underwater video fish target intelligent tracking and recognition system is designed.The recognition accuracy(Mean Average Precision,mAP)of the system for six kinds of fish videos is 98.50%and the tracking efficiency(Frames Per Second,FPS)is 46.The test data show that the system realizes the accurate recognition and real-time tracking of underwater fish video targets,and provides intelligent support for the scientization and digitization of aquaculture.