Safety Helmet Detection And Identification Based On Improved Faster R-CNN

Wearing safety helmet is one of the effective methods to prevent the head injury of operators on the construction site.However,most of the current helmet wearing detection algorithms have many shortcomings,such as difficult to detect occluded targets,low recognition accuracy of small targets,poor adaptability of complex scenes and so on.Therefore,according to the characteristics of the construction site environment and video streaming,this article divides the target into five categories of people wearing red,yellow,white,and blue helmets and those who do not wear helmets according to the color,and an improved Faster R-CNN algorithm for helmet detection and identification is designed to detect the wearing status of the helmet and and confirm the identity according to the color to judge whether it meets the construction requirements.Aiming at the problem of poor detection effect of Faster R-CNN caused by too low target resolution in the monitoring screen,feature fusion and multi-scale detection methods are used in this paper to improve the network structure.In this model,the deep features of strong semantics in VGG16 are fused with the shallow features of high resolution by using the additive fusion function,and then input to RPN layer by layer for multi-scale detection,so that the small target features can be preserved to the stage of classification and location,so as to improve the detection accuracy.To solve the problem of high loss caused by the imbalance of hard,easy,positive and negative samples in the data set,OHEM mechanism is introduced to Faster R-CNN after feature fusion.The OHEM can mine difficult samples with large losses,gather them into training set batches with only difficult samples,and train them pertinently,which can enhance the resolution of the model to the background and reduce the rate of missed detection.In order to further improve the recall rate and generalization of the model,this paper improves the non-maximum suppression(NMS)algorithm and proposes the Crowded-NMS algorithm.The improved algorithm adopts the Gauss weighted penalty function,sets the double threshold,and optimizes the original two segment function into three segments,so as to solve the problem of target missing detection and target frame redundancy of high-density population.After the above three sequence improvements,the helmet-identity Model of this paper is constructed.Because there is no public data set,this paper built a "Factory" data set which contains the monitoring video pictures of chemical plants,which is used for network training and testing.The experimental results show that the accuracy of Faster R-CNN in "Factory" data set is 66.7%,and the accuracy of improved helmet-identity Model is 94.2%.By taking the surveillance video as the model input,the algorithm realizes the monitoring and alarm of two kinds of violations in the surveillance,which are not wearing safety helmet and the operation beyond the boundary.The test results show that the model has low missing and error detection rates and strong robustness when applied to video detection.