Deep Learning-Based Construction Site Safety Protection Equipment Detection

Currently,the video monitoring analysis system of construction sites still relies on manual supervision.However,infrequent inspections by inspectors,long feedback time and a lack of safety awareness all lead to frequent safety accidents.Therefore,it is imperative to realize an intelligent and real-time construction site video analysis system,and it is also an important means for construction enterprises to develop intelligent safety management.For this topic,this article mainly studies the intelligent algorithm design of whether the worker wears safety protective equipment(helmets,etc.).The main contents of this thesis are as follows.1)Data Collection and Labeling: At present,in the development of deep learning-based applications,large amounts of data are needed for training.Therefore,this article collects many pedestrians and helmets image,cleans and labels the unlabeled data,and integrates all label data into the same storage form.2)Model Design and Training: In order to make real-time analysis of video data in a surveillance video system,the one-stage object detection algorithm YOLOV3(You Only Look Once version 3)and Center Net(Center Point-based Object Detection)based on Convolutional Neural Network(CNN)are selected as the detection models.Aiming at the problem of partially labeled in the collected dataset(some datasets only labeled the helmet),the network structure and loss function of the algorithms are redesigned by adding new labels and knowledge distillation.Experiments are designed to evaluate the performance of the improved algorithm.3)Model Evaluation and Optimization: In order to reduce the computing pressure of the server and the bandwidth of video data transmission,this article designs a lightweight safety protection equipment detection model focusing on embedded devices by combining Mobile Net V2(Efficient CNN for Mobile and Embedded Vision Application)and Center Net.The precision and speed of the lightweight model are optimized by redesigning the feature fusion module,so that the model can achieve a running speed of more than 40 FPS(Frames Per Second)in Jetson TX2.When the threshold of intersection over union(Io U)between predicted box and ground truth is set to 0.5,the average precision of pedestrians and helmets can reach 77.5% and 86.0%,respectively.The intelligent model designed has the advantages of fast running speed and high precision,which can well detect workers and helmets in actual scenes.Moreover,the model requires less computing resources,and can be conveniently deployed to embedded devices.