Research On Detection Algorithm Of Helmet Wearing Based On Deep Learning

Most of the safety accidents in oil and gas stations are related to staff violations,and wearing helmets improperly is the most common violation.Therefore,it is very important to automatically detect helmets wearing with object detection technology.The current helmet detection algorithms have problems such as poor robustness in complex scenes and low detection accuracy of small targets.Aiming at the above problems,this paper proposes two safety helmet detection methods with low error and high robustness.The main research contents are as follows:Aiming at the problem of accuracy degradation in the helmet detection algorithm caused by factors such as inconsistent light,small targets and local occlusion in complex oil and gas background,a helmet detection method that based on optimized residual block and posture prior information is proposed.Firstly,a dual-branch network model is designed,and the pose key point information predicted by the human pose estimation model is used as a guided feature expression to fuse with the YOLOv3 target detection branch to obtain more accurate detection results.Secondly,an optimized residual attention network structure is constructed based on the feedforward network,which combines the optimized residual structure derived by the NAG algorithm formula with the improved mixed attention module to enhance the network expression ability and suppress the interference of irrelevant features.Multi-branch dilated convolutional blocks are introduced into the prediction branches of three scales to expand the network receptive field to obtain richer global features.Lastly,taking pose as a prior constraint,an improved non-maximum suppression named PA-NMS is proposed to further suppress false detections and reduce missed detections.Aiming at the redundant calculation problem of similar results in the detection method based on dense prior,this article uses Sparse R-CNN network with preset sparse proposals combined with the enhanced FPN feature extraction network,and proposes a helmet detection network with an enhanced fusion of global and local features.Sparse RCNN can obtain Ro I coordinates through a small number of proposals,reducing redundant information and speeding up detection.The enhanced FPN feature extraction network can integrate richer contextual information in high-level feature maps to improve the feature extraction capability of feature pyramids.Meanwhile,for the problem of unbalanced samples in the public helmet dataset,the data augmentation method can be used for expansion.The helmet wearing detection system is designed and developed using the PyQt5 platform.The system calls and displays the results of various network models,calculates the prediction results from the input data,and provides a visual detection interface of different methods.Finally,a large number of experimental results on GDUT-HWD,SHWD,Hard Hat Workers-v2 datasets and the self-made oil and gas station dataset show that,compared with other methods,the helmet detection method proposed in this article is superior in precision and recall rate,etc.In the actual scene,the error of the helmet target detection task is lower,and the generalization effect is better.