Risk Target Identification And Tracking Of Transportation Construction Sites Based On Computer Vision

Transportation infrastructure is the lifeline of economic and social development of countries and regions.During the construction of transportation infrastructure,a large number of complex resources such as people,machines,materials and substances would be gathered in a limited space during a short period of time,and the complex interaction of construction resources with time and space constraints makes the safety management of construction sites extremely challenging.The state recognition of construction site risk targets can provide effective support for construction managers to obtain dynamic information about the site,thus ensuring the normal operation of the construction site and the safety of construction personnel.Current methods based on traditional contact sensing approaches cannot effectively calculate the effective and real spatio-temporal states of risk targets,and also suffer from the deficiency of high cost and low efficiency.To address the above limitations,this paper considers the characteristics of spatially "large dispersion and small aggregation" and time-varying changes,and gradually investigates computer vision-based risk target identification and tracking methods for transportation construction sites from sparse to dense distributed,from rigid to flexible,from spatial domain to spatio-temporal domain,and from 2D to 3D.The main research contents are as follows.(1)For the rapid identification of the location and status of sparsely distributed construction workers in spatially large-scale construction sites,a two-level construction worker information identification method based on image stitching technology was proposed.Firstly,the GPS and pose information from the UAV image capture was used to pre-align the construction site images,and a feature extraction and matching technique based on feature area filtering was established,the optimal stitching method was used to eliminate the ghost effect,after that the full-field image of the construction site was obtained.A classification network was employed to filter the worker areas according to the sparse distribution characteristics of the construction workers,and the location and status of workers were identified by using a detection network.(2)To address the problem of accurate recognition of densely distributed construction vehicles at construction sites,a dense construction vehicle information identification method based on orientation-aware bounding box(OABB)and feature fusion was proposed.A single-stage detection network with integrated OABB proposal and regression was established,and the feature fusion module was used to fuse the features from multiple levels to further improve the identification accuracy.The method was validated using various types of construction vehicle data from representative real construction sites.(3)For efficient state identification of articulated construction machines with higher working risk in construction sites,a lightweight network-based pose estimation method for articulated construction machines was proposed.Firstly,the pose of the construction machine was parameterized,the key node coordinates were processed using a twodimensional Gaussian distribution to generate heat maps,the pose estimation network was lightweighted at the network level,the module level and the channel level,and the L2function-based network loss function was constructed.Finally,the method was validated using an annotated dataset containing a variety of complex construction site backgrounds.(4)To address the problem of complete acquisition of spatio-temporal information of construction vehicles in construction sites,a 2D motion tracking method of construction vehicles based on OABB was proposed.The features of OABB were used to describe the 2D translational and rotational motions to represent construction vehicle contours,a contour detection module was built to obtain construction vehicle contours in complex backgrounds,the Kalman filtering technique was used to correlate the temporal changes of the contour information,and the tracking ID management module was constructed for the control of the intersection over union index.Finally,two-dimensional motion information was obtained from real construction vehicle videos.(5)For the acquisition of spatio-temporal information about working construction ships in construction sites,a 3D parameter tracking method of construction ships based on monocular vision was proposed.Using deep ship detection model trained by ship images collected from the Internet and the annotations,ship regions were located and the precise ship contours were obtained by morphological operations in transformed color space.Through constructing the known conditions of the site,3D spatio-temporal parameter calculation method was established based on the projection transformation model.The3 D parameters of the real construction ship were tracked to verify the feasibility of the proposed method.The research results of this paper will provide theoretical support for the state identification of various risk targets of transportation construction sites and provide data basis for the safety control of construction sites.