Research On Helmet Detection Algorithm Of Lightweight YOLO Combined With Pixelshuffle

With the development of Internet technology,computer vision related technology has penetrated into all aspects of social life,especially the target detection technology based on computer vision is extremely widely used.At present,China’s infrastructure projects are in full swing,construction sites are all over the country,but some workers lack of awareness of prevention,especially the lack of awareness of proper wearing of safety helmets,resulting in frequent accidents at construction sites.Therefore,it is particularly important to study the safety helmet detection algorithm suitable for the application of the site.Algorithms with excellent performance can not only reduce the losses of construction companies,but also ensure the safety of workers to a large extent.However,the interference factors of complex application scenarios have a great test of the speed and robustness of the algorithm,such as target occlusion,detection of small targets,dim light and other interference factors.In order to solve such problems,this paper conducts the following research based on the YOLOv4 network model:（1）In view of the problem of information loss during the upsampling process of YOLOv4,this paper uses pixel convolution（Pixel Shuffle）to improve the upsampling branch in the YOLOv4 network.A new upsampling module was constructed using subpixel convolution to replace the traditional interpolated upsampling to alleviate the problem of semantic information loss;secondly,for the problem that the parameters of the YOLOv4 network were more redundant,the Double-fire module was built in series between two Fire modules to replace the 5× CBL convolutional group,and the Ghost convolution was introduced.Based on this,this paper proposes a PFG-YOLO network model combined with subpixel convolution.（2）In order to further reduce the network model,this paper constructs a bottleneck structure that can replace the CSP structure based on the Ghost convolution,and uses the bottleneck structure to reconstruct the feature extraction structure Ghost Net of YOLOv4,which successfully replaces CSPDark Net-53;for the problem of redundant parameters in YOLOv4,a 3×3CBL convolution of the Neck part of YOLOv4 is introduced with a depth of 3×3 deep separable convolution to replace the Neck part of YOLOv4.In combination with the lightweight network Ghost Net,this paper proposes a lightweight version of PFG-YOLO,G-PFG-YOLO.（3）Through theoretical analysis,on the processed hard hat detection dataset,this paper compares and verifies the improved algorithm.Experiments show that the detection accuracy of PFG-YOLO is increased by 2.1% on the basis of YOLOv4,the detection speed FPS is increased by 50%,and the model volume is reduced by 20%;compared with YOLOv4,the detection speed of G-PFG-YOLO is about 300%,and the model volume size is only 19% of YOLOv4,which can meet the requirements of real-time detection on portable devices.Finally,This paper proposes two network models for helmet detection: PFG-YOLO and GPFG-YOLO.PFG-YOLO is an improved version of the helmet detection for YOLOv4,and the GPFG-YOLO is a lighter version of PFG-YOLO,which is more suitable for use on some civilian embedded devices.