Safety Helmet Wearing Detection Based On Deep Learning And Saliency

As one of the most common and practical personal protective equipment,safety helmet can prevent and reduce the damage to people’s head caused by external hazards effectively.Therefore,the correct wearing of safety helmet is the top priority to ensure safe production.The traditional method of manual inspection by security personnel is time-consuming,laborintensive and inefficient.However,with the development of artificial intelligence technology,the use of deep learning technology to solve problems in actual production and life has become more and more mature.We can use object detection technology to achieve accurate detection and real-time monitoring of production personnel wearing safety helmets.Simply using the object detection technology to detect the wearing of the helmet in the production environment will encounter the following difficulties: the complex environment will cause false detection of the background target.And the continuous actions of the production personnel will cause missed detections in it.It also repeated alarms for the same target on the same shooting device.In order to improve the detection accuracy of helmet wearing and solve the above problems,this paper mainly does the following research work:(1)For the network structure of the detection task,this paper proposes an improved network structure based on YOLOv4-Tiny.First,replace the backbone network with the better GhostNet-Tiny to extract features more accurately in less time.Second,add the SPP structure to the output of the deepest layer,which can greatly increase the receptive field and separate the features of the context.Finally,the original convolution block in the FPN structure of this network is replaced with a depthwise separable convolution block,which greatly reduces the model parameters and reduces the model size.Experiments of the improved model in the helmet test dataset show that each of the above three methods can improve the detection accuracy of the original model,and the final overall effect is significantly improved.(2)Aiming at the problem of false detection in actual detection,this paper proposes a data false correction algorithm based on saliency detection.The saliency detection technology used an improved video saliency detection based on fully convolutional network in this paper.First,the algorithm is used to generate the saliency estimation map of each frame to distinguish the foreground human object and the background area.Then,the recognition box obtained by object detection is compared in the corresponding saliency estimation map.If a box is in the background area,it will be judged as a false detection,otherwise it is a correct detection.The method is applied in the test set of continuous frame pictures,and the results show that it can effectively eliminate the false detection object and improve the detection accuracy.(3)Aiming at the problem of missed detection in continuous frame detection,this paper proposes a missed detection monitoring algorithm based on object tracking.In this paper,the SORT object tracking technology is used to compare the estimated box generated by the tracker with the next frame.If the estimated box does not have a corresponding real box,and the position of the estimated box is in the saliency estimation map’s foreground object area,it will be judged that there is a missed target at this position.Applying this technique in the same continuous frame test dataset as in the previous step,the results show that the number of missing objects can be effectively identified and the accuracy of the final result can be improved.