Research On Construction Violation Behavior Detection From Monitoring Perspective

Violations in the industrial construction environment may lead to huge economic losses and even life-threatening.Adhering to the operating rules mainly relies on the consciousness of workers and supervision of managers,but manual supervision is inefficient,easy to miss and cannot monitor every employe uninterruptedly.Therefore,the use of monitoring equipment to achieve intelligent detection of violations is of great significance.However,currently there is no public construction violation dataset towards monitoring perspective.Only few training data are available,which can’t cover the problems caused by changes in lighting,viewing angles and scenes in video surveillance.On the other hand,some construction violations depend on the interrelationship between people,objects and construction scenes.For exmaple,towards the action of whether to wear goggles when welding,existing behavior detection methods are sometimes difficult to detect.Aiming at these problems,this thesis has launched a series of researches on construction violation detection methods towards monitoring perspective.Based on the comprehensive consideration of model complexity and model speed,this thesis will focus on the application of object detection in violation detection based on monitoring perspective.First of all,aiming at the problem that there is currently no relevant public construction violation dataset,we constructe a monitoring based construction violation detection dataset,named VOMP.By manually observing the video,manually intercepting the video segment of the violation,manually labeling the violation,we generate labels of the violation dataset,and organize it into a standard object detection format.By comparing the difference between the monitoring based construction violation detection dataset and the typical object detection dataset,the characteristics and difficulties of theconstruction violation detection dataset are analyzed,and we also compared the experimental results of mainstream object detection algorithms on this dataset.Secondly,aiming at the problem of the small number of training samples and the sample imbalance in the connstruction violation detection dataset,this thesis proposes a pair-wised similarity data augmentation method for violation detection.If the two pictures are successfullt paired,the paste confidence map is calculated,and the area with higher confidence is randomly selected to exchange the foreground and background of the two pictures to generate a new picture.In order to weaken the edge effect when pasting,this thesis further uses the generative adversarial network to render the pasted image and make the synthetic image as realistic as possible.The pairing similarity data augmentation method for violation detection increases the sample diversity and significantly improves the accuracy on the VOMP dataset.Finally,in some construction violation scenes,there is often a correlation between people,operators,and construction scenes.In order to effectively detect such behaviors,context information is required.At the same time,in order to meet the speed requirements in practical applications,this thesis proposes an efficient method for detecting violations based on contextual information.The traditional single-stage network is improved by adding a context information branch.This branch is only used during training,so it does not hurt the network test speed and significantly improves the accuracy on the VOMP dataset.