基于DeepSORT算法的鱼道过鱼种类识别和计数研究  

Species identification and counting of fishes in fishway based on Deepsort algorithm

作　　者：吴必朗 柳春娜[1,2,3] 姜付仁 WU Bilang;LIU Chunna;JIANG Furen(China Institute of Water Resources and Hydropower Research,Beijing 100038,China;State Key Laboratory of Simulation and Regulation of Water Cycle in River Basin,Beijing 100038,China;Key Laboratory of River Basin Digital Twinning of the Ministry of Water Resources,Beijing 100038,China)

机构地区：[1]中国水利水电科学研究院,北京100038 [2]流域水循环模拟与调控国家重点实验室,北京100038 [3]水利部数字孪生流域重点实验室,北京100038

出　　处：《水利水电技术（中英文）》2022年第9期152-162,共11页Water Resources and Hydropower Engineering

基　　金：中国水科院“五大人才”计划专项项目(SD0145B032021);国家自然科学基金项目(51809291)。

摘　　要：为了解决鱼道过鱼监测中鱼类种类识别和计数问题,基于深度学习技术,以异齿裂腹鱼、拉萨裂腹鱼、拉萨裸裂尻和尖裸鲤为目标鱼类,在Y江某水电站鱼道开展过鱼种类识别和计数实验。首先,将过鱼视频制作成数据集,评估Faster R-CNN、YOLOv4和YOLOv5s 3种目标检测算法的性能。其次对DeepSORT算法中检测部分做出改进:将目标检测算法由Faster R-CNN替换为YOLOv5s,用于检测视频中的鱼类。然后采用卡尔曼滤波算法和匈牙利匹配算法对检测到的鱼类进行跟踪预测和最优匹配,通过对每一条目标鱼类分配唯一的ID,在视频中部划分虚拟“检测线”的形式,对过线的鱼类实现计数。结果表明:(1)Faster R-CNN、YOLOv4和YOLOv5s的鱼类平均检测准确率分别为80.85%、82.85%和86.6%。其中,YOLOv5s的准确率最高,异齿裂腹鱼、拉萨裂腹鱼、拉萨裸裂尻和尖裸鲤的种类识别准确率分别为94.19%、90.22%、84.57%和77.41%。(2)以YOLOv5s为检测器的DeepSORT算法鱼类计数平均准确率为71%,异齿裂腹鱼、拉萨裂腹鱼、拉萨裸裂尻和尖裸鲤计数准确率分别为80.7%、66.2%、64.8%和72.4%。研究结果能够为实现鱼道过鱼监测的自动化、智能化提供新方法,为鱼道的运行管理提供决策参考依据。To solve the problem of fish species identification and counting in fishway monitoring, we used deep learning technology to carry out the fish passage species identification and counting experiments. Our data set includes 4 target fish species collected at the Hydropower Station on the Y River basin such as Schizothoracus disodontii, Schizothoracus waltoni Regan, Schizothoracus younhusbandi Regan, and Oxygymnocypris stewartia. Firstly, we divide the data set into training set and verification set to evaluate the performance of three target detection algorithms(Faster R-CNN, YOLOv4 and YOLOv5 s). Secondly, we use YOLOv5 s algorithm as a detector to improve the detection part of DeepSORT.And the target detection algorithm is replaced by Faster R-CNN to YOLOv5 s to detect fish in video. Kalman filter algorithm and Hungarian matching algorithm will carry out tracking prediction and optimal matching for the detected fish. The fish crossing the line can be counted by giving a unique ID to the detected fish and setting a virtual “detection line” in the middle of the video. The results show that the average precision of Faster R-CNN, YOLOv4 and YOLOv5 s were 80.85%, 82.85% and 86.6%, respectively. According to the YOLOv5 s algorithm, the accuracy precision of identifying species of Schizothoracus disodontii, Schizothoracus waltoni Regan, Schizothoracus younhusbandi Regan,and Oxygymnocypris stewartia are 94.19%, 90.22%, 84.57% and 77.41% respectively.The mean accuracy of DeepSORT algorithm for fish counting was 71%, and the accuracy of Schizothoracus disodontii, Schizothoracus waltoni Regan, Schizothoracus younhusbandi Regan were 80.7%, 66.2%, 64.8% and 72.4%, respectively. This study provides a new methodology for realizing the automation and intelligence of fish passage monitoring and provides a reference basis for operation management of fish passage.

关 键 词：DeepSORT YOLOv5 鱼道过鱼种类识别 鱼道过鱼计数 深度学习 

分 类 号：TV213.4[水利工程—水文学及水资源] S956.3[农业科学—水产养殖] S932.4[农业科学—水产科学]