Construction Site Safety Protection Detection System Based On Deep Learning

With the development of the economy,the building industry is developing rapidly.And the detection of wearing helmets and safety belts at the construction site is becoming more and more important.On the one hand,many construction sites detect the safety equipment by manual monitoring,and the statement of workers' safety equipment cannot be obtained in time,lead to an increasing of the possibility of accident.On the other hand,some construction sites have intelligent video detection systems for the wearing of safety equipment detection with low recognition accuracy,delayed real-time video transmission and bad real-time performance.In order to solving these problems,this dissertation has designed a deep learning-based safety equipment detection system.Firstly,this dissertation introduces the related techniques about the proposed detection system,mainly about the image processing and object detection.And then,for the disadvantages of the traditional machine learning algorithm such as the sensitivity to light,angle and images' resolution,an improved Retina Net based safety equipment detection algorithm is proposed in this dissertation.In addition,this dissertation also proposed a method to increase the robustness of the model by adding Gaussian noise to the training set.Finally,by training the weights of the Retina Net model,this dissertation applies the algorithm and the Image AI video detection library to the JETSON Tx2 edge computer to realize an efficient real-time safety equipment detection system.The work and innovation of this dissertation are as follows:(1)By obtaining real-time monitoring video to be detected,we can upload it to the server side of the laboratory to ensure remote control and concurrent requirements of the system.And then,we use the FFmpeg to acquire and split the monitoring video to be detected on the server side of the laboratory to obtain real-time monitoring pictures.According to the training video,a number of training pictures are obtained,and each training picture is subjected to calibration of whether the worker wears safety equipment.In the training pictures,depend on whether the workers wear the safety equipment,the training video is converted into thousands of frames by FFmpeg software,and then calibrated by Label Img,the visual image labeling tool,to generate an XML file with the PASCAL VOC format.70% of the calibrated training pictures are used as the training set,while 30% of the calibrated training pictures are used as the test set.(2)In this dissertation,for the environment problems at the construction site such as much dust,uneven illumination distribution,low light intensity and low contrast,an image preprocessing method base on adaptive contrast histogram equalization algorithm is adopted.In addition,the Retina Net algorithm is adopted to solve the problems of traditional detection methods,such as susceptible to external factors,cumbersome detection steps,unbalanced positive and negative samples in the original one-stage detection algorithm.The experimental results show that the improved Retina Net algorithm is better than the original Retina Net algorithm and the Faster R-CNN algorithm,which can meet the actual detection requirements.(3)For the low real-time performance,slow detection speed and complex network deployment of the current intelligent video remote monitoring system,this dissertation applies the high detection speed Image AI to the real-time JETSON Tx2 edge computer for intelligent detection of safety equipment at the construction site.Compared with remote video surveillance,the combination of Image AI and JETSON Tx2 reduces the time of video transmission,improves the detection speed,reduces the cost of construction company's deployment of video surveillance network,improves the portability of detection tools,which is a feasible choice to intelligently monitors in the future construction industry.