Safety Helmet Wearing Status Detection Based On Computer Vision

Safety helmet,as a tool to protect the head of constructor,plays an extremely important role in the personal safety protection of workers.The existing helmet wearing status detection is mainly based on manual monitoring.Manual monitoring requires monitoring personnel to keep on duty at all times,however,manual monitoring is prone to oversight,there is a situation of missing inspection.Safety helmet wearing status detection based on computer vision can replace manual monitoring.Detecting the wearing of safety helmet in different construction scenarios has become a research hot spot in artificial intelligence,pattern recognition and other fields,especially in computer vision field.Aiming at the complexity of detecting the safety helmet wearing status,this paper proposes two detection methods in the process of image feature description,feature extraction,and object detection.The specific research work and innovations of this paper include the following two points:1.Propose an Improved Boosted Random Ferns Algorithm(IBRFs)and apply it to the detection of safety helmet wearing status.Aiming at the problem that helmet wearing status detection is easy to fail in complex architectural scenes,this paper takes random ferns algorithm as the basis.Firstly,the image features are extracted by Histogram of Oriented Gradient(HOG)to form the feature domain space of the image.Secondly,random ferns were constructed by random binary test in feature domain space.Then,the position points and parameters with significant difference between object and background were selected on random ferns to construct weak classifiers with certain recognition ability.Finally,the Real Ada Boost algorithm is used to construct a strong classifier for image classification and prediction.Since the construction of classifier in Real Ada Boost algorithm is only a simple weighted combination of multiple weak classifiers and does not play the role of weak classifiers well,an improved enhanced random ferns algorithm is proposed in this paper.The weighted coefficient method is used to select multiple weak classifiers with strong discriminant characteristics to build a strong classifier.In the real-time helmet detection task,the detection accuracy of the IBRFS model proposed in this paper reaches 92.74%,which is about 9.5% higher than that of the BRFS model,and greatly improves the classification accuracy of the classifier.In a normal environment,the detection speed of 8 f/ s can be achieved to meet the requirements of real-time detection.2.The ResNet50-SSD(ResNet50-Single Shot Multi-box Detector)network is constructed and applied in the safety helmet wearing status detection.In order to solve the problem of low accuracy of small-scale object detection in real-time scenes,this paper proposes a method of helmet wearing status detection based on ResNet50-SSD network.The model is based on the SSD(Single Shot Multi-box Detector)network.Firstly,the basic network VGG-16 is replaced with ResNet-50,and the structure of the first residual network of conv4＿x in ResNet-50 is modified to make the size of the eigenmatrix of the first output constant;Secondly,a series of additional layers are added after conv4＿x and the Batch Normalization(BN)layer is introduced to speed up the convergence of the network.Thirdly,the objects of different sizes are predicted in different feature layers.Finally,the non-maximum suppression algorithm is used to filter out the objects with low probability to get the final detection result.On the international pubilc safety helmet dataset SHWD,the detection accuracy of the ResNet50-SSD model proposed in this paper reaches 80.4%,Compared with vgg16-ssd model,it improves by 2.23%,and the detection speed is 35 frames per second,which meets the requirements of real-time detection.The two algorithms proposed in this paper are used to test the wearing status of safety helmets in real-time scenes datasets and international public datasets respectively,and compared with other existing advanced object detection algorithms,it verifies the superiority of the proposed algorithm in detection accuracy and real-time performance.Through experimental verification,the proposed algorithm can not only ensure high detection accuracy,but also meet the requirements of real-time detection.It has strong generalization ability and robustness,and can adapt to real-time detection of safety helmet wearing status in complex environment.