Helmet Wearing Detection And Its Applications In Multiple Scenarios

In order to ensure safe life and production,people are required to wear safety helmet in multiple scenarios.If they don’t wear safety helmet,there will be hidden danger.Therefore,there is a demand for safety helmet wearing detection in multiple scenarios.However,at present,in the construction site,substation,transmission tower and other scenes,there are still many problems in the safety helmet wearing detection,such as multiple false positives in the complex background,serious omission in multiscene object detection,and the redundancy of training model parameters,which makes it difficult to implement the engineering.Therefore,we propose a method of helmet wearing detection in multiple scenarios based on the improved version of YOLOv3（Helmet-YOLOv3）.Towards the problem of large number of false positives in object detection under complex background,we use SVM to model the data,and use the existing data for accurate training,so it can more effectively obtain the position information of personnel and safety helmet,and has good applicability in multiple scenarios.When labeling data,the coordinate information of personnel and safety helmet can be obtained.Based on the coordinate information,a SVM discriminator can be trained.Using the discriminator,we can predict the relationship between the person and the helmet detected by the model.When the false alarm occurs in the model,we can reduce the false alarm rate based on the SVM object relation discriminator.Towards the serious problem of missed detection in object detection in multiple scenarios,we use the channel attention mechanism to explore the relationship between different channels in the yolv3 model,and adaptively recalibrate the weight value of channels.Without increasing the number of feature maps,we can learn global information more effectively to selectively enhance feature representation and suppress useless information,so as to improve the accuracy of helmet wearing detection in multiple scenarios.Towards the problems of large redundancy of training model parameters and slow reasoning speed,we apply channel pruning algorithm to YOLOv3 model in practical application.In the training process,L1 regularization is used as the penalty term of scaling factor in batch normalization（Batch Normalization,BN）layer,which makes scaling factor in some BN layers gradually become zero.Through this operation,unimportant channels can be identified in the helmet wearing detection task,so as to carry out subsequent model compression.In this way,we can get a compact YOLOv3 network with few parameters,small model and low computation.As there is no relevant public dataset,we collected images of construction sites,substations,streets,security and other scenes,and built a dataset-"helmet" for algorithm training,verification and testing.From the experimental point of view,the model proposed in this paper is compared with the general object detection model in mAP,inference time and model size,and then the ablation experiments of the three improved modules are carried out.From the experimental results,we can see that compared with the original YOLOv3,the mAP of Helmet-YOLOv3 model is improved from 90.1%to 98.1%,the reasoning time is reduced by 7 ms,and the model size is reduced by 200 M.Finally,the model is deployed in the actual multi-scene monitoring environment.From the actual detection results,it can be seen that the performance of the proposed model is good,and it can carry out real-time and accurate helmet wearing detection in multi-scene environment.