改进YOLO5Face的小鼠行为实时分析方法研究  

Research on real⁃time analysis method of mouse behavior by improved YOLO5Face

作　　者：胡春海[1] 姜昊 刘斌[1] HU Chunhai;JIANG Hao;LIU Bin(Key Laboratory of Measurement Technology and Instrumentation of Hebei Province,Yanshan University,Qinhuangdao,Hebei 066004,China)

机构地区：[1]燕山大学河北省测试计量技术及仪器重点实验室,河北秦皇岛066004

出　　处：《燕山大学学报》2023年第4期359-369,共11页Journal of Yanshan University

基　　金：河北省自然科学基金资助项目(F2019203511)。

摘　　要：传统的动物行为分析方法大部分是采取离线的形式,不能做到实时分析。为了解决此问题,本文提出了一种改进YOLO5Face的小鼠行为实时分析方法。本方法分为两个步骤:首先是小鼠关键点实时检测,然后是小鼠行为实时识别。针对小鼠关键点实时检测,在深度学习网络YOLO5Face的基础上改进:新增了一个更小的检测头来检测更小尺度的物体;主干网络中加入YOLOv8的C2f模块,让模型获得了更加丰富的梯度流信息,大大缩短了训练时间,提高了关键点检测精度;引入GSConv和Slim-neck,减轻模型的复杂度同时提升精度。结果表明:模型对鼻尖、左耳、右耳、尾基关键点检测的平均PCK指标达到了97.5%,推理速度为79 f/s,精度和实时帧率均高于DeepLabCut模型的性能。针对小鼠行为实时识别:利用上述改进的关键点检测模型获得小鼠关键点坐标,再将体态特征与运动特征相结合构造行为识别数据集,使用机器学习方法SVM进行行为分类。模型对梳洗、直立、静止、行走四种基本行为的平均识别准确率达到了91.93%。将关键点检测代码与行为识别代码拼接,整个代码运行的实时帧率可以达到35 f/s。Most of the traditional animal behavior analysis methods take the form of offline,can not achieve real⁃time analysis.In order to solve this problem,a real⁃time analysis method of mouse behavior is proposed by improving YOLO5Face.In this paper,two steps are proposed:firstly,real⁃time detection of mouse key points,and then real⁃time recognition of mouse behavior.For the real⁃time detection of key points in mice,some improvements are made on the basis of deep learning network YOLO5Face.Firstly,a smaller detection head is added to detect objects of a smaller scale.Secondly,The C2f module of YOLOv8 is added to the backbone network,which enables the model to obtain more abundant gradient flow information,greatly shortens the training time,and improves the detection accuracy of key points.Thirdly,The complexity and accuracy of the model are both enhanced by the introduction of GSConv and Slim⁃neck.The results show that the average PCK index of the model for nose tip,left ear,right ear and tail base key points reaches 97.5%,and the reasoning speed is 79 frames per second.Both the accuracy and real⁃time frame rate are higher than the performance of DeepLabCut model.Real⁃time recognition of mouse behaviors:the improved key point detection model is used to obtain the key point coordinates of mice,and then the posture features and motion features are combined to construct the behavior recognition data set,and the machine learning technique SVM is employed to classify the behaviors.The average recognition accuracy of the model for four basic behaviors,namely grooming,rearing,resting and running,reaches 91.93%.By combining the key point detection code with the behavior recognition code,the real⁃time frame rate of the whole code can reach 35 frames per second.

关 键 词：小鼠行为识别 关键点检测 实时性 改进YOLO5Face 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]