Safety Recognition Of Personnel Behavior In Thermal Power Plants Based On YOLO

In the production work area of the thermal power plant,the unsafe behavior of the staff often leads to accidents,if an accident occurs,it will cause bad consequences to the employees and enterprises.In order to prevent the occurrence of safety accidents,it is of great significance to study the real-time detection of unsafe behavior of personnel in the production work area.Nowadays,with the rapid development of computer vision technology,You Only Look Once,as an open source detection algorithm,has obvious advantages in speed and accuracy compared with other algorithms in the same period.Combined with YOLO target detection algorithm,this paper makes an in-depth study on the recognition of unsafe behaviors such as not wearing helmet and smoking.The main work of this paper is as follows:A YOLO-Gaussian helmet wearing recognition algorithm based on YOLO algorithm is designed,which aims at the problems of low recall rate,inaccurate location and poor accuracy of small target detection in YOLO v3 algorithm.In this paper,the algorithm is improved in two aspects of network structure and loss function without losing the detection speed.First,the Gaussian model is introduced to predict the location uncertainty of the position parameters of the model bounding box.Increase the output of position parameters,on the basis of which the location loss function is reconstructed;Secondly,the dense connection structure is introduced into the feature extraction backbone network to strengthen the feature transfer between network layers,enhance feature reuse and alleviate the problem of gradient disappearance in the process of network training.In this paper,the verification is carried out on the selfmade data set of helmet-wearing.The experimental results show that without losing the detection speed,m AP0.5 increases by 2.54% and m AP0.75 is increased by 32.68%respectively.A lightweight real-time smoking behavior detection algorithm based on YOLO v3-tiny algorithm is designed.Due to the high hardware requirements of YOLO v3 algorithm,YOLO v3-tiny model is often chosen,but this model has the problem of low detection accuracy.In this paper,the priori frame of smoking data set is designed by k-means clustering algorithm to improve the recall rate of the algorithm,and the spatial pyramid pooling structure is introduced to enhance the feature extraction of small targets to improve the detection accuracy.The experimental verification on the self-made smoking data set shows that the m AP0.5 is increased by 24.4%,which proves the effectiveness of the algorithm.