Research On Safety State Monitoring Technology Of Construction Workers Based On Computer Vision

Civil engineering construction requires a large number of construction workers to participate,and the construction process is relatively high risk..Head and neck injuries of construction workers are the most serious and even fatal in all construction accidents,and the number of accidents that die from head trauma in construction safety accidents per year remains high.Wearing a safety helmet can effectively reduce the physical impact damage in a risk accident.However,in construction,due to the weak awareness of workers' safety,construction personnel often fail to wear safety helmet in accordance with production specifications,resulting in great potential safety hazards.Therefore,it is of great significance for the automatic detection of whether the construction personnel wear helmets.This paper presents a helmet detection method based on computer vision human pose estimation.Main contents are included as follows:Firstly,a method of safety helmet wear detection for construction personnel based on deep learning is proposedTarget detection algorithm is applied to detect whether the constructors wear safety helmet or not.The object detection algorithm is studied.The construction site image data set is collected and produced,and the object detection model is established and trained.Through the object detection system,the locations of the two kinds of target objects in the surveillance video image sequence can be detected.Secondly,a method of safety helmet wear detection for construction personnel based on improved YOLOv3 model is proposed.Because there are many small objects in the video sequence,in order to further improve the recognition accuracy of small helmets in the video image,the model structure is optimized and adjusted,and the network is optimized in three aspects: target dimension clustering,multi-scale detection and dense connection.Firstly,by using dimension clustering method,according to the size characteristics of constructors and safety helmets in the training set,the statistical rules are found,and nine kind of anchor boxes are clustered to replace the anchor boxes in the original model.Subsequently,in view of the small object of the safety helmet in the perspective of the construction site,a fourth scale detection part is added,and 12 anchor boxes are clustered,which are allocated to four scales in order of size,and a four-scale detection model is obtained.Finally,using DenseNet network model for reference,the cross-scale prediction part of the network structure adopts the dense connection mechanism,which connects the four scale layers of the feature fusion part densely,and maintains the size of the feature map of different scale layers consistent with the size of the feature map of the predicted output scale through the up-sampling process.The detection results of the optimized models were compared by experiments.Finally,a safety helmet wearing detection method based on OpenPost algorithm is proposed.The human pose estimation is applied to the scene of automatic detection of whether the constructors wear safety helmet or not.An automatic detection system integrating human pose estimation and object detection is designed.Firstly,the human frame output from the target detection model is matched with the human pose output from the corresponding human pose estimation model.Then the human frame is traversed.According to the spatial relationship,the safety helmet is paired with the human body,and the correct wearing of the safety helmet is judged according to the spatial position relationship between the safety helmet and the central point of the human face.