| With the continuous advancement of urbanization and public infrastructure construction,the number of construction projects continues to increase.However,due to the lack of safety awareness of site workers,a large number of safety accidents caused by safety violations occur at the construction sites every year.Methods such as conducting regular safety training and arranging safety inspectors to conduct on-site inspections can only have a temporary effect on reducing the rate of safety accidents on the construction site,and cannot always supervise the safety violations of the construction site.In recent years,the use of surveillance cameras widely installed on the construction site as the visual terminal,and the use of computer vision technology to provide safety warnings on the violations of the workers on the construction site is a new hot-spot research,and it is quite important for ensuring the life safety of workers and the production safety of the construction site.Regarding the construction site safety early warning,this thesis mainly completes two sub-tasks: one is to detect whether the workers on the construction site are wearing hardhats,and the other is to detect whether the workers on the construction site are in the hazardous work area of heavy equipments.For the task of hardhat wearing detection,the existing pedestrian detection-based method has low detection accuracy due to occlusion problem;the face detection-based method also has the problem of scene limitation,and the generalization ability of the model is not strong.Therefore,this thesis proposes a detection method that directly detects whether a human head is wearing a hardhat.The main work of the thesis includes the following points: 1.the heads wearing hardhats and those not wearing hardhats are classified and labeled as different wholes,and a larger and more diverse set of hardhat wearing detection data set has been constructed;2.by replacing the backbone network,adding DCN(Deformable Convolutional Networks)or SAC(Switchable Atrous Convolution)to improve the Cascade R-CNN object detection algorithm,a multi-scale hardhat wearing detection network structure that can adapt to the object scale transformation has been obtained.After training and testing on the constructed hardhat wearing data set,the calculation results of performance indicators show that the hardhat wearing detection model achieves a maximum 92% value of AP(Average Precision)in the detection of head wearing hardhat,and a maximum 94% value of AP in the detection of head without wearing hardhat.Compared with the current best performance of hardhat wearing detection model,the value of m AP(mean Average Precision)has increased by 7.01% to 92.9%;3.according to the size of object,the degree of crowding and occlusion,and illumination,a series of visual comparison tests have been performed on 7 scenes.The test results show that the hardhat wearing detection model in this thesis has achieved great performances in various scenarios,which proves that the model is less affected by factors such as crowding,occlusion,illumination,and scale conversion.The reliability of the multi-scale hardhat wearing detection network structure proposed in this thesis is also verified.For the task of detecting whether workers on the construction site are in the hazardous work area of heavy equipments,the main work of this thesis includes the following points: 1.construct a data set composed of pedestrians and heavy equipments with Mask information,and train on the Cascade Mask R-CNN object detection algorithm.The trained model is used to extract the Masks of workers and heavy equipments;2.propose a new set of pixel focal length self-calibration algorithm based on pedestrian detection technology,and verify the effectiveness of the algorithm on the KITTI data set;3.proposed a set of depth reconstruction algorithm for pedestrians and a set of filtering and denoising algorithm for heavy equipments;4.according to the conversion principle,the 2D objects on the RGB images are converted into the 3D objects in the point cloud space for processing,and a set of point cloud space safety warning algorithm is proposed.By testing on the constructed annotated video data set,the algorithm’s early warning accuracy reaches 95.99%.In addition,the point cloud space visualization test results show that the algorithm achieves great early warning effects on both video data sets,which further validates the effectiveness of the safety early warning algorithm in point cloud space. |
PDF Full Text Request