基于计算机视觉的鱼类低氧胁迫行为检测与跟踪算法  

Detection and Tracking Algorithm of Fish Hypoxia Stress Behavior Based on Computer Vision

作　　者：李道亮[1,2] 姜国旗 杨建安 白羽 谢琰 王承国[3] LI Daoliang;JIANG Guoqi;YANG Jian'an;BAI Yu;XIE Yan;WANG Chengguo(National Innovation Center for Digital Fishery,China Agricultural University,Beijing 100083,China;College of Information and Electrical Engineering,China Agricultural University,Beijing 100083,China;Yantai Institute,China Agricultural University,Yantai 264670,China)

机构地区：[1]中国农业大学国家数字渔业创新中心,北京100083 [2]中国农业大学信息与电气工程学院,北京100083 [3]中国农业大学烟台研究院,烟台264670

出　　处：《农业机械学报》2023年第10期399-406,共8页Transactions of the Chinese Society for Agricultural Machinery

基　　金：国家自然科学基金项目(62076244);山东省科技厅项目(2022LYXZ012)。

摘　　要：为了能准确检测、跟踪加州鲈鱼因水中溶解氧含量低产生的胁迫行为,本文构建了一种改进的YOLO v5与DeepSORT组合网络算法。在算法方面提出2个改进方案:在原YOLO v5的Backbone和Neck中分别加入2个基于移位窗口的自注意力Swin Transformer模块,提升了网络对目标特征信息的提取能力,以此提升原模型的检测效果;采用Warmup和Cosine Annealing结合的学习率策略,使多目标跟踪算法DeepSORT前期收敛速度更快、更稳定。实验结果表明,在目标检测方面,相对于原YOLO v5,改进的YOLO v5的mAP@0.5、mAP@0.5:0.95和召回率分别提升1.9、1.3、0.8个百分点,在不完全遮挡情况下,改进的算法表现出更好的检测效果。在目标跟踪方面,DeepSORT算法的MOTA、MOTP和IDF1分别提升4.0、0.7、10.7个百分点,并且加州鲈鱼在遮挡前后的ID切换频率得到明显抑制。改进的YOLO v5与DeepSORT跟踪算法更适合于检测、跟踪加州鲈鱼的低氧胁迫行为,能够为加州鲈鱼的养殖提供技术支持。In order to accurately detect and track the stress behavior of micropterus salmoides due to low dissolved oxygen content in water,an improved YOLO v5 and DeepSORT combined network algorithm was constructed.In terms of algorithm,two improvement schemes were proposed:two self-attention Swin Transformer modules based on shifted windows were added to the Backbone and Neck of the original YOLO v5,which improved the network's ability to extract target feature information,thereby improving the detection effect of the original model;the learning rate strategy combined with Warmup and Cosine Annealing made the convergence speed of the multi-target tracking algorithm DeepSORT faster and more stable in the early stage.The experimental results showed that in terms of target detection,compared with the original YOLO v5,the mAP@0.5,mAP@0.5:0.95 and recall rate of the improved YOLO v5 were increased by 1.9,1.3 and 0.8 percentage points,respectively.In the case of incomplete occlusion,the improved algorithm could show better detection results.In terms of target tracking,the MOTA,MOTP,and IDF1 of the DeepSORT algorithm were increased by 4.0,0.7 and 10.7 percentage points respectively,and the ID switching frequency of micropterus salmoides before and after occlusion was significantly suppressed.The improved YOLO v5 and DeepSORT tracking algorithms were more suitable for detecting and tracking the hypoxic stress behavior of micropterus salmoides,and can provide technical support for the breeding of micropterus salmoides.

关 键 词：加州鲈鱼 胁迫行为 目标检测 多目标跟踪 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]