| Wearing a safety helmet when entering construction sites,production and other places can reduce the injuries caused by falling objects,collisions,etc.,and protect personal and property safety,but for various reasons,there will always be people who do not cooperate with these requirements.In the past,administrators often through on-site supervision,view monitoring or the use of sensors to determine whether the staff wear safety helmets,and give early warnings,this way is time-consuming and labor-intensive,and there will be poor supervision or accidental negligence,resulting in some avoidable accidents from time to time,so the design of a set of intelligent automatic identification management system,the personnel who do not wear safety helmets to identify,prompt,block,record,etc.It is very necessary,the core of the detection system is the detection algorithm.The existing detection algorithms have the problem of missing detection and misdetection when detecting targets wearing safety helmets in complex environments(poor light,dense personnel,many obstacles,long distance,etc.).In view of this,this paper optimizes and studies the safety helmet wearing detection algorithms for construction site workers in complex environments based on deep learning that can be embedded in detection systems,as follows:(1)The development history and research status of safety helmet detection methods are introduced,some problems existing in mainstream algorithms are analyzed,the basic concepts of object detection algorithms based on deep learning,such as convolution,pooling,activation,etc.,are introduced,and then two-stage and single-stage detection algorithms are analyzed,mainly introducing three types of YOLO series algorithms.(2)Aiming at the problems of occlusion,illumination,low image clarity and small target,a helmet detection algorithm(SC-YOLOv5)with detection layer and attention module is proposed.Based on YOLOv5s(5.0),a small target detection layer is built to extract more detailed outline,color,shape and other features of the target to enhance the semantics and spatial information richness of the feature map.The channel attention module and spatial coordinate attention module are inserted to generate channel and spatial coordinate attention feature maps to improve the accuracy of model detection.In order to reduce the detection impact caused by dataset labeling errors,the online labeling software was used to label the safety helmet and portrait target on the image by itself,and a 7600-sheet safety helmet dataset was produced.After training tests and actual detection,it is proved that the SC-YOLOv5 algorithm improves the accuracy by 1.4%,and the performance is better than the original YOLO series detection algorithm and some mainstream detection algorithms,and has certain excellence.(3)The shortcomings of the first improved algorithm in detection rate are analyzed,and a detection algorithm based on lightweight convolution is designed.Based on YOLOv5s(6.0),by modifying part of the convolution to lightweight convolution,the low-definition target information is enriched,and the detection time is controlled while replacing the small target detection layer to improve the detection performance of the algorithm.Update the loss function,introduce angles,optimize parameters,improve the bounding box loss function,improve detection performance and make the model converge faster.Embed the CA module to optimize the weight of the feature map and extract useful features.Experiments show that the improved algorithm improves the mean average accuracy by 1.1% and the number of processing frames by 9.5%,which is only half of the loss of the first improved algorithm,and the improved algorithm has significantly better performance than the original algorithm under the influence of detecting low-definition targets and various interference terms.(4)Design a set of monitoring sampling to detection processing to record archiving detection system,with two improved algorithms as the core,analyze the function and feasibility of the system,design the system architecture,database and functional modules,and design a simple,safe and easy-to-use helmet wearing detection system.In summary,this paper studies and improves the current mainstream target detection algorithms,proposes two optimization algorithms,improves the detection performance of the algorithm for the safety helmet wearing detection of operators,and designs an automatic detection system in order to adapt to the actual detection task.Using the algorithm proposed in this paper and any feasible detection system,it has certain practical value to detect the safety helmet wearing target of complex construction site. |
