Research On Visual Inspection Of Safety Belt Wearing For High-altitude Operations

Infrastructure facilities have become an important pillar of our country’s economy,in which there are many high-altitude workers.Many risks potentially exist in the complex and changeable scenes for high-altitude operations.Main reasons of the injuries and deaths for high-altitude operations workers include no-wearing or irregular wearing of safety belts,which will possibly cause heavy casualties.As a precaution,safety belt plays a very important role in protecting high-altitude workers from accidents.However,due to some reasons,safety belts has not been strictly used in many situations.In order to early alert the workers in high-altitude operations,some studies based on image processing have been provided to automatically detect whether the workers are wearing safety belts.In this thesis,a deep-learning-based belt detection for the workers in high-altitude operations is proposed to deal with the limitations of the existing visual studies in accuracy and real time.1.This thesis proposes a visual detection method of safety belts for the workers in high-altitude operations.First,a safety belt dataset for high-altitude operations is established.Then,a safety belt detection model is proposed based on a convolutional neural network to detect whether the workers wear safety belts in high-altitude operations.The thesis mainly includes two jobs as following.In order to effectively detect whether a worker over there wears a safety belt,the image features are extracted by a multi-scale convolutional network.Here,safety belt detection is regarded as a two-category classification problem.Dilated convolutions and depth separable convolutions are employed in its feature extraction to improve the overall receptive fields and to enhance the convergence of the network,which can tend the network to a balance between accuracy and real time.The proposed progressive attention region mechanism focuses on global and local information existed in the images,which can further improve the detection accuracy for small targets.2.The proposed deep network is an entirely end-to-end network,which does not require manual feature design and selection.It provides some modifications in feature extraction and fusion,which can improve the feature expression ability of the network.A loss function is proposed to effectively reduce the difference between the real value and the predicted value,and to enhance the network convergence.Also,it is verified by the experiment.Experimental results show that the proposed safety belt detection method performs well under various high-altitude operation situations.When the size of the input image is 608×608 pixels,it can achieve an average accuracy of 80.22%.