Research On Visual Detection For Personnel Safety Protection On Construction Sites

The construction industry,as one of the backbones of China’s economy,incurs a huge demand for construction workers.However,construction sites are of complexity and variation,containing a number of potential hazards,such as falling objects,collapse of working platform and mechanical injuries,which can cause casualties to workers.The main causes of construction casualties include brain injuries caused by falling from a height and collisions.As a preventive step,the usage of personal protective equipment(PPE)for workers plays an essential role in protecting construction individuals from accidents(e.g.hardhats and harness).However,the usage of hardhats and harness is not strictly enforced among workers due to all kinds of reasons in some cases.In order to conduct pre-warning for those workers who have not used their PPEs,some studies have been presented on detecting non-hardhat-use and noneharness-use of workers based on image processing and deep learning.To break through the potential limitations of the generality and efficiency of existing studies,this thesis proposes some visual detection algorithms for personnel safety protection on construction sites based on deep learning,including hardhat detection and harness detection.For hardhat detection,a new hardhat wearing detection benchmark dataset is constructed for various construction site scenarios,involving complex backgrounds,variations in weather and light,and individual occlusions.Then,the hardhat wearing detection problem is defined as a 5-classes object detection problem(hardhat of four colors and non-hardhat-use).Also,a hardhat wearing detection framework is proposed based on convolutional neural network(CNN).To overcome the difficulty of the small-scale hardhat detection,the reverse progressive attention(RPA)is proposed to fuse features from different layers of different scales,which discriminately generates a new feature pyramid for detection.Experimental results demonstrate that the proposed method is effective under all kinds of on-site conditions,which can achieve 83.89% mean average precision(m AP)with the input size 512×512.For harness detection,the harness detection problem is considered under resourcerestricted conditions and a lightweight harness detection framework is proposed by combining dilated convolutions and depthwise separable convolutions.In order to meet the real-time requirements of processing surveillance video in the video surveillance platform in construction sites,an integration module is proposed as the basic unit to construct the structure of backbone feature extractor,which is constructed by dilated convolutions and depthwise separable convolutions.The principle of this module is to conduct the dilated convolution with different dilation rates to each channel of the feature map,which can reduce parameter sizes while expanding its receptive field.Experimental results show that the lightweight harness detection model can run at a speed of more than 25 frames per second(FPS)and achieve a much better trade-off between computing resources and detection accuracy.The proposed hardhat-wearing and harness-wearing detection algorithms will provide new solutions to the safety monitoring of personnel on the smart construction site.