Design And Implementation Of A Deep Learning-Based System For Detecting Abnormal Behavior At Construction Sites

With the rapid development of China’s economy,the process of construction industrialization is accelerating,but safety accidents caused by irregular operations of construction personnel still occur frequently every year.Traditional abnormal behavior control measures mainly rely on manual methods,but due to insufficient strength and poor effect,a large number of images captured by surveillance cameras cannot be effectively processed.Therefore,it is necessary to introduce more intelligent and automated solutions to improve security control.efficiency.In recent years,the rapid development of intelligent information technology has provided effective support for the automatic detection of abnormal behavior.In particular,computer vision technology represented by deep learning does not need to attach equipment to workers and has little impact on their operations.At the same time,it can realize efficient and accurate processing of massive image data,and is widely used in construction site scenes.However,the current related research mainly focuses on the analysis of production efficiency,and rarely discusses the research on abnormal behavior detection,and most of them are limited to the detection of a single abnormal behavior.At the same time,abnormal behavior detection has high requirements for the recognition of elements such as scene characteristics and real-time action,which also brings challenges to its detection accuracy.In view of this,this paper proposes a method for detecting abnormal behavior on the construction site based on computer vision technology,which comprehensively considers system,real-time and accuracy to improve the detection effect.Based on the research status at home and abroad,this paper divides the abnormal behavior of the construction site into two types: static behavior and dynamic behavior.Among them,the static behavior detection is aimed at the detection of the wearing behavior of the construction workers and the fireworks scene on the construction site,and the dynamic behavior detection is aimed at the detection of the construction workers’ falling and climbing over the railings.This paper uses surveillance video to detect the above two types of abnormal behaviors.The main work is as follows:(1)Use target detection algorithm to realize static behavior detection.Aiming at the problem that the existing open source data sets are not suitable for the actual site environment,this paper cooperates with the relevant engineering construction units of Country Garden to obtain video data by using the monitoring video of the construction site,and expand the original data through data enhancement methods such as adjusting brightness,mirror flipping,and left and right rotation.Increased by 5 times,and further proposed a lightweight target detection model YOLOv5s-Mobile Net V3 based on YOLOv5 s improvement.After experimental data verification,the YOLOv5 sMobile Net V3 model has faster reasoning speed while maintaining the accuracy basically unchanged.The size of the improved model is only 7.4MB,the number of parameters is reduced by 50% compared with the original model,and the FPS is increased by 3.6 times.The YOLOv5s-Mobile Net V3 model has excellent detection performance,and has a good performance in static behavior detection,which can meet the actual construction site application requirements.(2)Using HRNet and Bi LSTM to realize dynamic behavior detection.Firstly,a selfbuilt construction site behavior data set was used,and the HRNet model was used to extract the key points of the skeleton and convert it into a CSV format data set.In order to reduce the computational complexity of the model,this paper extracts valid motion data samples from all bone point data through a keyframe-based method,and performs data preprocessing operations,such as deleting redundant joint points and supplementing missing joint points.After feature extraction,a dynamic behavior detection model is built using SVM,Random Forest,XGBoost,LSTM and Bi LSTM.Through experimental comparison,it is found that the Bi LSTM network model shows high accuracy in the selfbuilt data set,and can realize real-time detection of dynamic behavior.(3)Design and implementation of abnormal behavior detection system on construction site.Based on the above research content,this paper designs and develops a construction site abnormal behavior detection system.After a detailed demand analysis of the system,the system is divided into a multi-channel video processing module,a static behavior detection module,a dynamic behavior detection module and an offline training module.Four modules are introduced in detail.Finally,complete the realization of the system and demonstrate the functions and interfaces.