Research And Application Of Fall Behavior Detection Based On Improved YOLOv5

Fall detection technology is of great significance in building intelligence,as it can improve the safety of living environments,reduce the risk of death and potential risks after recovery caused by falls,and provide effective monitoring tools for medical staff,family and public place managers.With the popularization of surveillance devices such as cameras and the maturity of computer vision technology,video based fall detection methods have become a better choice.This method has lower invasiveness towards the elderly and lower cost in actual deployment.Therefore,this article proposes a more efficient and lightweight YOLOv5 improved model for fall detection.The main research work is as follows:(1)Aiming at the problems of YOLOv5 model in fall detection,an improved method of fall detection model based on YOLOv5 was proposed.Firstly,by introducing the CBAM attention mechanism into the feature extraction network of YOLOv5 and improving the original bounding box loss function,the detection accuracy of the model is improved,and the probability of missing or false detection of the model is reduced.Then,Bi FPN is used as the feature fusion structure in the Neck section of the model,and a detection layer that can detect smaller targets is added by introducing shallow features into the feature fusion structure,This improves the detection ability of the model for small targets.Through the above improvements,this article has obtained a new model for the fall detection task: YOLOv5-CCB.The experimental results show that the improved model’s m AP improves by 4% compared to the original YOLOv5,and the overall performance of the model has also been greatly improved.(2)Aiming at the real-time requirements of fall detection in current practical scenarios,a lightweight method for fall detection model based on improved YOLOv5 is proposed.This method uses YOLOv5-CCB as the initial model,first using the lightweight network Shuffle Net V2 as the feature extraction network of the backbone,then performing channel pruning operations on it to further reduce the volume of the model,and finally using knowledge distillation to optimize the pruned model to improve the detection accuracy of the lightweight model.Through the above lightweight processing,this paper obtains a lightweight fall detection model based on improved YOLOv5: YOLOv5-SCCB.The experimental results show that the weight of the model is only 51% of the original model,and the inference speed is increased by 1.4 times.Compared with other mainstream target detection algorithms and fall detection algorithms,the fall detection model in this paper has certain advantages in detection accuracy and detection speed.(3)A fall behavior detection system is designed and implemented by integrating the improved method and lightweight method of the fall detection model based on YOLOv5.In the design and implementation of the system,an embedded industrial computer with high reliability and good compatibility was selected as the hardware platform of the system.Python language was used to implement the system’s functions such as rapid selection of fall detection models,adjustment of relevant thresholds,statistics of the number of falls.The Py Qt5 framework was used to build a graphical interface for the fall behavior detection system,improving the convenience of system operation.Figure [52] Table [14] Reference [60]...