Research On Quantitative Gait Analysis Method Of Experimental Mice Based On YOLOv5

Automated and accurate quantitative analysis of rodent gait characteristics plays an important role in kinematic,neurological and aerospace medicine research,and also plays an important part as an aid in studies related to gait analysis in ground-based simulated weightlessness conditions.Although methods for rodent gait studies have evolved considerably over the years,they all have their limitations.Existing research methods for gait analysis in rodents have typically used traditional image processing algorithms for target detection.They are mainly sensitive to noise,slow,inaccurate,have low generalization and require manual error correction.They are not well adapted to different colours and species of rats and have specific hardware requirements.To address the shortcomings of existing gait analysis methods,this paper proposes a gait analysis system for experimental mice based on target detection network using C57BL/6J experimental mice as the research object,and designs experiments to verify its reliability,while also carrying out gait analysis of experimental mice under ground simulated weightlessness conditions,the main work is as follows.(1)Optimization of the YOLOv5 network model for experimental mouse paw print recognition.In this paper,the YOLOv5 model is used as the main framework for recognition of this dataset,combining wavelet analysis with the YOLOv5 framework in parallel to enhance the network’s ability to extract features and improve the network’s recognition of high-frequency small targets;optimising the structure of the PANet so that features converge to the output of detecting small targets;adopting a channel pruning algorithm to design the model in a lightweight manner,trimming unimportant channels,and removing the output of detecting large targets according to the characteristics of the actual dataset.And according to the characteristics of the actual data set,the output side of detecting large targets is removed,thus reducing the number of parameters and improving the operation speed.The experimental results show that the improved network has a 0.8%improvement over the pre-m AP@[0.5,0.95] with almost the same computational speed,and the pruned network has a 1.4% reduction in m AP@[0.5,0.95] but a doubling of the computational speed.(2)An improved quantitative gait analysis method based on YOLOv5 mouse gait(Mouse YOLO)is proposed.Firstly,we designed and built an experimental mouse gait acquisition bench,collected and produced gait data sets with a total of 8530 sample labels;then we quantified and analysed the gait characteristics of mice,including the analysis of gait characteristics parameters and the visualisation of gait;finally,we compared Mouse YOLO with Mouse Walker and the manually calculated method to assess the validity of the model.Overall Mouse YOLO was slightly better than Mouse Walker and comparable in accuracy to the manual calculation,but the running time Mouse Walker was30 times longer than Mouse YOLO and the manual calculation took several hours more.(3)Experiments were designed to simulate weightlessness model in experimental mice.Application case studies of Mouse YOLO were conducted.We designed the mouse simulated weightlessness model experiments and control group experiments,and used Mouse YOLO to detect significant changes in gait parameters in mice under simulated weightlessness relative to the control group.By visualizing the results of gait,the differences between the simulated weightlessness group and the comparison group could also be compared more clearly.This set of experiments shows that Mouse YOLO can well detect the unique gait differences of the simulated weightlessness model.