| With the continuous improvement of economic development and traffic infrastructure construction,people are more convenient access to the same time also face a lot of trouble caused by the frequent occurrence of traffic accidents.However,traditional traffic accident investigation methods are time-consuming and low in accuracy,which can hardly meet the needs of modern traffic management.Therefore,traffic accident scene investigation methods based on 3D reconstruction technology are gradually emerging as a new solution for traffic accident investigation,which can truly restore the scene of the accident.3D reconstruction technology has been widely used in many fields,including face recognition,medical image reconstruction,robot path planning,building damage identification and disaster prediction.However,the traditional methods have great limitations,while the deep learning-based methods have better reconstruction effects but require higher hardware requirements.Therefore,this paper introduces the 3D reconstruction technology into the traffic accident site investigation,and proposes the Improved Panch Match Net and Structure From Motion(IPMSFM)algorithm.The algorithm cross-integrates the traditional Structure From Motion(SFM)algorithm with the improved PMN(Patch Match Net)deep learning network,it gives full play to the advantages of traditional methods and deep learning methods,and can efficiently and accurately reconstruct the three-dimensional model of the traffic accident scene.On this basis,a traffic accident scene 3D reconstruction system is designed and developed,which can realize the 3D reconstruction process and display the 3D model and other traffic information comprehensively.The research content of this paper is divided into the following aspects:1.To address the problem that the 3D reconstruction using deep learning can obtain better reconstruction effect but with low efficiency and high equipment cost,the sparse 3D model of the scene is obtained by using SFM,and the sparse depth map is obtained by interpolating the generated sparse point projections,which replaces the random initialization of the original PMN,thus reducing the computational complexity and memory overhead,and further optimizing the loss function of PMN to obtain higher quality This can improve the accuracy of the reconstructed model,speed up the reconstruction speed,and reduce the training time of the model.2.IPMSFM is trained using a large-scale,open-source multi-view 3D reconstruction DTU dataset and tested on a dataset simulating traffic accident scene scenes,and analyzed in comparison with MVS(Multiple View Stereo)method and PMN.The results show that the reconstruction time of IPMSFM is reduced by 13,215 s compared with MVS and 8,782 s compared with PMN,and the mean error of IPMSFM is reduced by 37 cm and 18 cm compared with MVS and PMN,and the median error is reduced by 28 cm and 12 cm compared with MVS and PMN,respectively.The experimental results show that IPMSFM outperforms MVS and PMN in terms of completeness,reconstruction efficiency,and accuracy of reconstructed model.3.The traffic accident scene survey requirements were studied and the traffic accident scene was simulated accordingly.An efficient and feasible photography scheme was selected through experimental tests,and the simulated traffic accident scene was photographed by a UAV to obtain a high-quality traffic accident scene dataset,which was subsequently processed and analyzed by using IPMSFM to generate a highly realistic 3D traffic accident scene reconstruction model.4.A traffic accident scene 3D reconstruction system was designed and developed,containing two parts: software for realizing 3D reconstruction process functions and a large visualization screen of traffic data for displaying 3D models and comprehensive traffic information.The system provides traffic police and road administration staff with a tool for traffic accident scene review and analysis at any time in order to accurately determine the cause of accidents. |
PDF Full Text Request