Real-time Wearing Of Helmets In Subway Construction Scenarios Monitoring Method Research

In recent years,there have been frequent safety accidents in subway construction sites.Irregular safety helmets worn by construction workers and the inability to accurately grasp the location of personnel are the main causes of accidents.Therefore,a real-time monitoring method for helmet wearing including personnel tracking is proposed for construction sites.It can effectively protect the personal safety of construction personnel.This thesis studies a method for real-time monitoring of helmet wearing in subway construction scenes.The monitoring video of the subway construction site environment is used as the research object,and deep learning is used to detect the people who need to be detected and their helmet wearing conditions,and perform Tracking,counting and warning to achieve the purpose of intelligent monitoring.Aiming at the problem that traditional image detection methods have poor detection results when applied to subway construction environments,this thesis designs a helmet wearing detection method based on the improved Tiny-YOLOv3 algorithm.This method does not require manual design features and is real-time.Good and other characteristics,through algorithm improvement and comparative performance evaluation,it is obtained that the improved detection method is close to the Tiny-YOLOv3 algorithm in detection speed,and has a significant improvement in the accuracy and recall rate of small targets.Aiming at problems such as missed detection of small target persons,the feature extraction part of the original network structure is analyzed,and the feature extraction of small target persons is improved by introducing the residual network structure into the feature extraction CBL module.Aiming at the problem of omissions in the selection of a priori frames for some features,by introducing the K-means++clustering algorithm,6-size anchor points are designed as the prior frame screening of the model,and the training speed is accelerated and the small target recall rate is improved.the result of.Aiming at the problems of the model’s redundant prediction box and prediction box labeling errors,the original loss function evaluation is replaced by improving the distance and intersection ratio of the loss function,and the loss rate and misjudgment rate of the prediction box are reduced.Aiming at the problems of repeated counting and error reporting in the monitoring system,this thesis adds a personnel tracking algorithm to the monitoring system to ensure the continuity of video frame monitoring.The DeepSORT tracking algorithm adopted by the tracking algorithm is improved from the SORT algorithm.Under the condition of ensuring the real-time performance of the algorithm,the problem of the loss of the target after the original algorithm is occluded is improved,which is in line with the complex application scenarios of the construction site.Finally,this thesis combines the hardware and software design on site,and integrates the software application of the helmet detection algorithm and personnel tracking functions.The experimental results show that compared with the original Tiny-YOLOv3 algorithm,the YOLO-hat detection algorithm proposed in this thesis has improved recognition accuracy by 4.6%,recall by 3.9%,and average precision by 4.1%,which proves the improvement of detection algorithm The feasibility and effectiveness of the monitoring system after combining the YOLO-hat detection algorithm and the personnel tracking algorithm meets the needs of on-site monitoring in practical applications.