Research On Multi-scale Object Detection Algorithm And Application In Construction Safety Monitoring

With the development of urbanization,a large number of infrastructure such as subways need to be built,and the safety of construction sites has attracted more and more attention,wearing safety helmets is an important guarantee for life and production safety in various engineering environments,and supervising workers wearing safety helmets has also become a necessary measure.The existing manual supervision method of wearing hard hats is far behind the current information-based and intelligent production mode under Industry 4.0.How to achieve intelligent helmet wearing detection,and how to accurately remind individuals to become a hot issue to be solved after detecting that the helmet is not worn.Based on the cutting-edge deep learning object detection technology and gait recognition technology,through the research and improvement of related algorithms,this paper proposes a helmet wearing detection and identification scheme for personnel without wearing a hard hat in the field of computer vision.In order to realize the detection of helmet wearing and the identification of personnel without helmets in construction scenarios,this paper proposes an integrated framework for detection and identification,which is composed of two modules: object detection and gait recognition.Through the comparative study of recent deep learning object detection models,the object detection module in this paper selects YOLO v5 as the baseline model,and the gait recognition module selects Gait Graph based on human skeleton extraction with many advantages as a baseline model.After the detection video is entered into the frame,it will first be unframed as a single frame picture,and then the helmet wearing detection will be carried out for each frame,and the human body boundary frame of the corresponding person will be speculated according to the detection frame of the unwearing hard hat,and then the human skeleton extraction will be carried out in the human body boundary frame,and finally the skeleton synthetic skeleton sequence extracted from each frame will be entered into the gait recognition network to confirm the identity of the person who is not wearing the helmet.At the same time,this paper also evaluates the target detection network YOLO v5 and gait recognition network Gait Graph to achieve better performance.In this paper,the attention mechanism SE layers introduced before the features of the YOLO v5 skeleton part are transmitted to the neck to avoid the loss of key information in the process of network transmission;when the feature scale fusion stitching is performed in the neck,the additional lateral connection from the lower layer of the skeleton part leads to the fourth detection scale,and the first three stitches are connected to expand the scale range of the model.It better detects small targets,and transfers the feature extraction ability pre-trained on large dataset MS COCO during training,which alleviates the problem of insufficient data set and improves the generalization ability of the network.For Gait Graph,a gait recognition network,this paper selects HRNet with stable performance and good performance for human skeleton extraction,and Res GCN,the basic module of Gait Graph network structure Analysis and research were carried out.By drawing on the CSP structure in the field of object detection,this paper improves the bottleneck structure of the spatial graph convolution and the spatiotemporal graph convolution that make up the Res GCN,which improves the learning ability of the network and reduces the memory consumption.Finally,the proposed hard hat detection algorithm and the improved gait recognition algorithm are tested on the public dataset,respectively,and the effectiveness of the improvement is proved.In order to test the detection and identification integration framework proposed in this paper,this paper uses the helmet dataset and gait recognition dataset to make a walking video of the person who does not wear a helmet under the construction scene,and the experiment proves that the detection and recognition integration framework proposed in this paper can determine the identity of the person who is not wearing a safety helmet.