| Through the study of animal behavior,we can explore their nervous system,psychological mechanism and response to drugs,usually using experimental mice as research objects.Most of the current animal behavior recognition methods are processed offline,because behavior recognition generally involves two steps of feature extraction and behavior recognition,which require a relatively long data processing time.Aiming at the poor real-time performance of the current animal behavior recognition methods,an improved YOLO5 Face real-time mouse behavior recognition method was designed and implemented.The main work of this paper is as follows:(1)Researched and implemented a real-time detection algorithm of mouse key points.Based on the deep learning algorithm YOLO5 Face,this paper improves and proposes a mouse key point detection model with both accuracy and real-time performance.In order to solve the problem that the mouse target is relatively small in the open field experiment scene,a smaller detection head is added to the head part of the network,which not only improves the detection accuracy of small targets but also increases the scale of feature fusion;in order to improve the characteristics of the entire backbone network Extraction ability,replace the C3 module in the backbone network with a more advanced C2 f module,which makes the model more accurate and converges faster;in order to reduce the complexity of the model and improve the model accuracy,introduce GSConv and Slim-neck in the network neck,and this structure can also make the attention module work better.In order to enhance the fusion between different channel features,the C3 module originally used in the detection head is replaced with the C3 SE module in the head part.The C3 SE module is a fusion of the C3 module and the SE attention module.The ablation experiment was carried out under the mouse key point detection data set produced in this paper,and the comparison experiment was carried out with Stacked Hourglass and the current excellent algorithm Deep Lab Cut for animal pose estimation.(2)A real-time recognition model of mouse behavior was constructed based on the combination of body characteristics and movement characteristics.This paper uses the dynamic time programming algorithm and ten machine learning classification algorithms to build a mouse behavior recognition model,and compares these algorithms on the mouse behavior recognition data set made in this paper,and verifies that they are in the experimental scenario of this paper.The support vector machine algorithm has the best classification effect.Based on the support vector machine algorithm,a real-time mouse behavior recognition model combining body features and motion features was built,and the algorithm difficulties and algorithm implementation process were analyzed in detail.The video image resolution of the input model is 640*640.From the image input model to realtime tracking of mouse key points to real-time behavior recognition,the real-time frame rate of the whole model can reach 35 frames per second.The mouse behavior proposed in this paper The recognition model has better real-time performance.Finally,the model is used to build a system interaction interface,which is convenient for users to interact without threshold.Users can freely choose to call the camera for real-time detection or to detect the video that has already been shot.The interactive interface of the system can realize model building,real-time tracking of key points,real-time recognition of behavior,and real-time drawing of trajectories. |
