Research And Implementation Of Fence Crossing Violation Detection Based On Deep Learning

In the safety management of the work site,the supervision of the fence crossing by nonconstruction personnel is essential.But at present,there are many problems in construction site,such as wide working area,which lead to low efficiency of manual supervision.As an important research hotspot in the field of computer vision,video-based human behavior detection technology has a broad application prospect in the field of public security monitoring.When violations occur in the monitoring images,it can be detected timely and accurately,greatly avoiding possible accidents.On the other hand,human behavior detection often results in poor detection effect due to factors such as large movement similarity and fuzzy boundary,so how to detect fence crossing violations more efficiently and accurately is an urgent problem to be solved to ensure the smooth progress of construction operations.In this context,this paper proposes a deep neural network based Fence Crossing Action Detector(FCAD)algorithm,which combines two-dimensional space features and threedimensional time sequence features to solve the positioning and classification of actions respectively.In order to solve the problem of inaccurate positioning of movement during fence crossing detection,FCAD used the improved Darknet-19 network to extract 2D spatial features from key frames,and realized the fusion of local features and global features through the SPP(Spatial Pyramid Pooling)module,which enriched the expression ability of the final feature map and improved the positioning effect;With a small staff resolution leads to fuzzy information,and then influence the performance of classification problem,FCAD personnel Detector(Person Detector)processing to expand resolution in original sequence,the staff more detailed features description,3d convolution with weighted sharing strategy 3D-Res Net-101 respectively to extract sequence of personnel(Actor Feature)and the characteristics of Scene(Scene Feature),improve the precision of classification of action;In order to further improve the detection effect,FCAD uses channel splicing to integrate the features of each part,and introduces the attention mechanism to attach weight,so that the network can pay more attention to important information;Finally,the detection results at the frame level are connected to form the final action pipeline,and the action pipeline with the highest comprehensive score is regarded as the optimal path to truly complete the detection of the fence crossing movement from the perspective of video;According to the size ratio of the personnel when the action occurs,the width to height ratio of the prior frame is predetermined by means of dimensional clustering,which makes the prediction frame more consistent with the real frame and speeds up the convergence speed of the model.Experiments show that the FCAD algorithm proposed in this paper is effective on the test data set.The detection speed can reach 31 FPS on 16 frames of input clip,and the Frame-AP(Io U=0.5)can reach 89.2%,which shows good performance in practical application.