Research On Video Intelligent Detection Method For Helmet

In the process of intelligent detection of helmet in oil production plant,the environmental background is complex,the object is occluded and the proportion of the object in the image is too small,which lead to the insufficient extraction of the object feature of helmet.In this thesis,an object detection model of helmet based on optimized dense connection and Gaussian modeling is proposed.First home-made helmet datasets,and the anchor box size of datasets is reconstructed by an improved clustering method.Then dense block and attention mechanism are introduced in the backbone feature extraction network,they can retain the object feature map information of low-layer network to a great extent and make the object feature information of hardhat get more weight.A deformable convolution module is added to the multi-scale detection branch to expand the size of the convolution receptive field and increase the detection probability of small objects.At the same time,because YOLOv4 model can only output the determined value of the position parameter of the detection box,it does not know whether the detection box is reliable,the Gaussian model is introduced into the algorithm to model the location parameters of the detection box,and the unreliable detection boxes are excluded,which further improves the performance of the network model.In order to ensure that the algorithm identified the helmet in the head position of human body,this thesis designs wear detection branch,the main method is to identify the staff image by YOLO-tiny and then get the head image by the head region localization algorithm,and carry out the helmet detection.The results of the two branches are selectively fused and reasoning to get the panoramic result.Experimental results show that the accuracy and recall of the proposed algorithm are significantly improved.The test results on helmet datasets show that the proposed algorithm has good detection performance.Aiming at the lower performance of lightweight network model than traditional target detection network,a lightweight helmet recognition model with enhanced feature fusion is proposed.Firstly,Ghost Net network is used to replace the original YOLOv4 model feature extraction network,which greatly reduces the cost of the model and ensures the performance of the model.Then,the CBAM attention module is optimized and improved,so that the feature map output by the attention module contains both spatial information and channel information with the coordinate information of the feature map.Moreover,the improved CBAM attention module is added to the Ghost Net network.The smooth Leaky-Relu activation function is deduced according to the method of deriving Swish activation function from Re LU activation function and applied to the model.Finally,based on PANet,a cross-layer feature fusion algorithm with weight is proposed.In this way,the network can more effectively fuse the feature maps of different scales.Through various experiments and tests on self-made datasets,the proposed helmet detection model has the same performance as the traditional object detection model under the condition of low parameters and floating point calculation,and greatly improves the reliability of the lightweight helmet detection algorithm.The research of helmet video intelligent detection method gives full play to the advantages of real-time monitoring system and massive resources,and provides scientific and technical support for oilfield safety production.