Automatic Identification Of Violatations Based On Smart Safety Helmets

With the development of artificial intelligence,the construction process of smart construction sites has been further promoted.In order to automatically detect relevant violations of construction site operators and prevent ineffective supervision due to manual inspections,this paper adopts deep learning to realize automatic monitoring of relevant personnel violations.Due to the fixed position of the camera on the construction site,there is a blind spot for vision,and the smart safety helmet can move with person to achieve targeted data collection.Therefore,this paper starts from the construction site data collected by smart safety helmets,and uses deep learning to automatically identify the two violations of on-site workers not wearing safety helmets and high-altitude scaffolding workers illegal wearing safety belts.The identification of violations in this paper is carried out around the key points of the human body.The related difficulties and work contents are as follows:(1)The detection algorithm of human key points and inclined target box is studied.In order to use the key points of the human body to identify subsequent violations,this paper adopts the center-point-based,bottom-up recognition algorithm model Center Net to identify the key points.At the same time,in order to adapt to the different horizontal shooting angles of smart safety helmets and realize the positioning of scaffolding area,this paper changes the prediction branch of Center Net and constructs the R-Center Net network model to locate the scaffolding area with different horizontal angles.On the basis of the above two algorithm models,a comparative experiment was carried out with different backbone networks,and the most suitable algorithm model was selected;(2)The algorithm of helmet wearing detection and climbing behavior recognition based on human key points is studied.The helmet wearing detection is not based on the traditional target box annotation,but uses the key points to locate the head area,and then uses PCA+SVM algorithm to predict the results.After narrowing the recognition range by positioning,the results show that this method has good recognition effect and detection speed.The identification of climbing behavior is mainly aimed at the intersection of the target box of the person and the bottom area of the scaffold.In order to overcome the influence of the background change of the frame before and after the data,the experiment compares the movement trajectories of some key points of person under different postures,finally get the best CSR threshold range for climbing behavior.Experiments show that this method has a good recognition effect,and at the same time makes good use of the key points of the human body,eliminating the time for additional image processing;(3)The identification algorithm of seat belt lanyard is studied.When the seat belt is in the illegal wearing state of low hanging height,the seat belt lanyard will appear below the knee of the operator.Therefore,in order to realize the recognition of small targets,this paper uses the Deep Lab V3+ algorithm model to detect the located seat belt lanyard area by means of semantic segmentation.In order to improve the identification effect,it provides data support for the alarm operation of subsequent violations.Based on the original model,this paper adds a feature fusion mechanism and an improved CBAM-f attention module.The ablation experiment results show that the improved model Deep Lab V3+-AC improves m IOU and F1 scores by 3.5% and 2.7% compared with the original model,and obtained the best recognition effect.Through the above work content,a novel and feasible reference idea is provided for the identification of relevant violations,and it also has certain reference significance for the follow-up in-depth research in this direction.