Research On Construction Machinery Target Detection And Recognition Based On Improved YOLOv5

As a labor-intensive industry,construction safety accidents occur from time to time,and a large number of construction safety accidents have caused irreparable human and material losses and negative social impacts.The complex and difficult construction environment of the construction site,the high turnover rate of construction workers,and the simultaneous operation of different construction machinery also increase the complexity of construction management and greatly increase safety risks.According to the statistics of the Ministry of Housing and Urban-Rural Development,the interaction between construction resources(such as construction workers and construction machinery)is one of the important factors leading to the production safety accidents of housing and municipal engineering in China.In recent years,with the improvement of computer hardware performance and the development of deep learning technology,visual target detection technology has been applied in various fields.Video surveillance at construction sites is used for dynamic observation of construction progress,site order and safety,however,traditional surveillance video requires manual intervention and analysis,which is less efficient and cannot meet the demand for real-time monitoring and alerting,therefore,intelligent analysis of construction site video surveillance is the trend.Among them,the target detection as the key technology of video surveillance system intelligence,to provide basic data support for safety management and decision-making.Therefore,in order to achieve accurate detection of construction machinery targets,this paper deeply investigates visual target detection technology based on deep learning theory and methods,and improves the YOLOv5 model in order to improve site safety management efficiency while reducing safety management costs,mainly including the following aspects of work.First,the construction machinery types are extended to establish a complete construction dataset.In response to the defects of the current public construction machinery dataset with uneven quality and single detection type,this paper increases both construction machinery types and image sample size,and establishes a homemade construction machinery image dataset with 22,315 images of 11 types of construction machinery by data enhancement techniques to provide a large dataset for the subsequent research of target detection algorithm.Second,the feature fusion network is improved to improve the detection accuracy of YOLOv5 model.To address the problems of background interference and object scale differences in construction scenes,this paper proposes an improved YOLOv5 model based on multi-scale feature fusion,replacing the feature pyramid(FPN)in the YOLOv5 network structure with a bi-directional feature pyramid network(Bi FPN)with learnable weights,learning the importance of different input features,and then fusing different input features with differentiation.Helps the network to find regions of interest in images with large area coverage,suppresses interference from complex backgrounds,and improves the overall accuracy of construction workers and construction machines.Third,the backbone network is optimized to accelerate the detection speed of YOLOv5 model.In response to the limitation of the hardware equipment of the actual construction scene,the deep learning model is difficult to deploy in the actual scene application,this chapter incorporates Rep VGG neural network in the backbone network of YOLOv5,uses multiple branch structures in model training,and uses the fused single branch structure model in model inference,which helps the model to reduce the number of parameters and make the weight file smaller in application inference,which not only can improve the computing speed,but also facilitate model deployment.Finally,experiments were conducted on the construction machinery dataset and compared with mainstream target detection algorithms.The experimental results show that the Rep-B-YOLOv5 model in this paper has a recognition accuracy of 89.7% and a recall rate of 82.2%,which are 3.1% and 5.1% higher than YOLOv5,respectively,proving the effectiveness of the model,which can better perform the detection of construction machinery targets and recognition tasks.The research in this paper proposes an algorithm optimization strategy based on the YOLOv5 model,which provides new ideas in the field of target detection;provides the establishment of a more complete construction machinery data set for the construction field,which is conducive to the further application research of target detection technology in construction sites;improves the speed and accuracy of construction machinery identification,which helps to improve the safety management of construction sites.