Research Of Object Detection And Risky Behaviour Recognition In Construction Scenes

With the continuous development of information technology,the construction industry is gradually moving towards modernization and informatization.Due to the level of industrial development,mechanized construction is currently unable to completely replace manual labor at this stage.The risky behaviour of building workers may lead to construction safety accidents,and cause damage to life or property.The existing technology uses manual inspection or wearing sensors to monitor the risky behaviour of building workers,which requires many human and material resources and it also affects the efficiency of workers.How to use artificial intelligence technology to detect risky behaviour of workers accurately has become a key research issue in the concept of "Smart Construction Sites",which is of great significance to improve the information management of safety risks at construction sites.Based on techniques related to object detection,pose estimation and graphical convolutional neural networks in computer vision,this dissertation focuses on the analysis of worker’s behaviour through skeletal information.To address the problem of difficulty in balancing model lightness and detection accuracy,a model combination of Yolox-s and Lightweight Open Pose is designed to extract bone nodes,and dense connection structures is used instead of the residual connection structures in a two-stream adaptive graph convolutional neural network.A new discrimination method based on the difference of the angular change rate between key points is proposed to address the issue of low classification accuracy of falling behavior using traditional threshold analysis methods.This dissertation selected two types of risky behaviors recognition tasks in construction for in-depth research: binary classification of falling behavior and multiclassification of ladder climbing behavior.The specific research works are as follows.1.In the binary classification task of detecting falling behavior,this study acquires bone nodes by inputting the end-to-end object detection results(human)into a bottom-up pose estimation network.The Yolox-s model optimized by CIOU loss function is used for the detection stage and the Lightweight Open Pose model is used for the pose estimation stage.The obtained information of the neck,left knee and right knee key points are input into the proposed algorithm for the classification of falling behaviour.To demonstrate the feasibility of this method,this study collected single-person falling behavior videos in laboratory scenes and multi-person falling behavior videos in construction scenes as two datasets and conducted experiments on both datasets for detection,pose estimation,and falling behavior recognition.The experimental results show that the model used in this study meets the requirement of model lightness while ensuring high detection accuracy.The overall behavior recognition accuracy rate of this method achieves 96.94% on the single-person dataset and 93.63% on the multi-person dataset.2.In the multi-classification task of ladder climbing behavior,this study inputs the preprocessed skeleton sequence information into the improved dual-stream adaptive graph convolutional neural network for training and obtains behavior classification results.The dual-stream adaptive graph convolutional neural network can process both bone node and bone length information for better spatial and temporal feature extraction.This study proposes to use global dense connection structures and local dense connection structures instead of residual connection structures in the original network,in order to reduce deep gradient vanishing and shallow feature reuse.To demonstrate the advantages of these two improvements,X-Sub and X-View experiments were conducted on the NTU60 RGB+D dataset.The results showed that both improved models were superior to the original model using residual connections in terms of recognition accuracy rate and model parameters.To test the practical effectiveness of these two improvements,this study collected ladder climbing behavior videos of workers in construction scenes as the dataset and trained multiple models for comparison after the same data preprocessing operation.The results show that the dual-stream adaptive graph convolutional neural network improved by global dense connection structures achieves the best classification effect in both mean average precision and mean average recall.Figure [23] Table [17] Reference [91]...