Lightweight Neural Network For Intelligent Construction Site Video Surveillance System

With the development of 5^th generation mobile networks,Internet of Things,artificial intelligence and so on,intelligent video analysis technology has ushered in rapid development.Different from the traditional video surveillance technology,intelligent video analysis technology can automatically analyze and recognize video image information without manual intervention.In construction sites,safety helmets are essential safety protection products for construction workers.Wearing safety helmets correctly is a guarantee for the safety of construction workers.However,at present,the detection of construction workers’helmets mainly relies on manual detection,which not only wastes a lot of labor costs and is inefficient.Using intelligent video analysis technology to replace manual detection,truly achieve information management of safe production.Aiming at the problem of helmet detection,traditional object detection methods are currently mainly used.Such methods are based on hand-designed features for identification,and the algorithm has poor robustness and low detection accuracy.With the development of deep learning,the accuracy of object detection has been greatly improved,but deep neural networks rely too much on hardware support,which is not conducive to the actual implementation of detection tasks.For this reason,this paper proposes intelligent construction site video surveillance system based on lightweight neural network.The work and results of this paper are as follows:（1）In view of the fact that there is no public safety helmet detection dataset suitable for this project,this paper collects video data from the construction site and constructs a safety helmet detection dataset through manual annotation.Through the use of geometric transformation and color transformation data augmentation methods,and the integration of the public dataset SHWD（Safety Helmet Wearing Detect Dataset）,the problem of too few effective targets for the helmet is solved.（2）The mainstream target detection algorithm YOLOv3（You Only Look Once）is used for helmet wearing detection,and the construction workers who are not wearing helmets are marked and warned.The Kalman filter combined with the Hungarian matching algorithm is used to track the walking trajectory of construction workers to prevent them from entering the dangerous area.（3）Aiming at the problem that the YOLOv3 algorithm has poor real-time performance and cannot be deployed at the edge,this paper optimizes the YOLOv3 neural network,replaces its main network framework,and uses depthwise separable convolution to replace the traditional convolution layer.Compress and quantify the trained neural network through Tensor RT,accelerate its inference runtime,and deploy it on the edge computing development board NVIDIA Jetson TX2.Different from the traditional server-centric centralized data processing method,the intelligent construction site video surveillance system based on lightweight neural network proposed in this paper takes edge computing technology as the core,adopts distributed data processing method,and directly processes video at the edge,which reduces the transmission of video data and the occupation of network resources.It reduces the cost of video surveillance network deployment by construction companies,makes up for the defects in surveillance,and truly achieves early warning,normal detection during the event,and standardized management after the event.