Research On Techniques Of Measuring Traffic Accident Scene Based On Computer Vision

Paper title: Study on Techniques of Measuring Traffic Accident Scene Based onComputer VisionMajor: Traffic Information Engineering and Control Department Advisor: Prof. Li JiangSome traffic accidents brought about some questions for society andtransportation environment .With the number of traffic accident increased, trafficcongestion formed badly affect traffic volume. How to reduce traffic accident sceneclearance time and quickly measure traffic accident scene is the hotspot question.Now, the tools for measuring the scene in our country is the tape, photograph issecondary used. The time is more and work efficiency is low. In addition, techniqueanalysis of traffic accident obligation needs to accomplish traffic accidentreconstruction and obtain some actual data. It is important for traffic accidentreconstruction to detect distant information from photograph by some newtechniques.Photography is used for traffic accident investigation by some scholars. Withthe development of computer vision, photos of accident scene are used to detectthree-dimension information which is required. The process from photos tothree-dimension information is named as computer vision, which consists of imagecollection, camera calibration, image matching, 3D reconstruction. The paperdescribe how to setup the calibration reference objects, some key questions such ascalibration reference point detection, camera calibration, 3D reconstruction areresearched.(1) Traffic accident had the effect on transportation, which is one of the maincauses resulting in traffic congestion. How to investigate accident scene quicklyand reduce clearance time are worthy of research. Accident investigationequipment in china and outside is analyzed, such as the tape, the laser scanner, thedigital photogrammetry system, combined total station and digital photographysystem. The condition that Photogrammetry used at traffic accident scene isdescribed. Now, some scholars always pay attention to reconstruct the points basedon stereo vision, which concludes the development of some aspects, for example,the number of cameras, the calibration reference objects, camera calibration, 3Dreconstruction, etc. Computer vision has the capacity of achieving 3D informationfrom the photos. The basic knowledge of computer vision is described, the linearcamera model and nonlinear camera model .The relations the world coordinatesystem, camera coordinate system and image coordinate system is discussed.(2) Detection of calibration reference point directly affects the result ofcamera calibration and 3-D reconstruction. Corners on the calibration referenceobject are seen as calibration reference points. How to detect quickly and preciselycalibration reference points is taken into account based on the analysis of thetechniques of Image processing such as edge detection and corner detection. Thealgorithm of corner detection based on Harris is brought to detect the referencepoints with better credibility, stabilization, practicality. Harris algorithm is thefeature point detection in computer vision with simple calculation and highprecision.Several special reference objects were designed to camera calibration such asthe planar pattern with the chessboard and the grids, 3D reference object. Aorientation sign is designed to detect some elements located at accident scene suchas vehicle, the brake trail , objects and road marking .The experiment resultsdemonstrate the algorithms is good to accomplish corner detection.(3) Camera calibration is a key step for 3D reconstruction from 2D images.Methods of Camera calibration is described with the introduction of linearcalibration and nonlinear calibration. The calibration reference object is developedfrom 3D poles and 3D objects to 2D plane. We proposed a method of cameracalibration based on 2D plane. The proposed procedure consists of a closed-formsolution, followed by a nonlinear refinement based on the maximum likelihoodcriterion. The procedure for two camera system is increased to obtain the locationrelation both two cameras on the basis of individual interior and exteriorparameters.The interesting factors affecting calibration error consists of the number ofphotos, the dimension of calibration reference objects, the distance from the object.We designed the experiment scheme to complete camera calibration for theconditions both one camera and two cameras. We gave a appropriate project fortraffic accident scene. The experiment results demonstrate that the method is goodfor camera calibration with low error .We think that calibration reference objectwith big dimension has effect on photos , while small dimension is bad for thereference point. We proposed the calibration object with 70 mm grid. While thenumber of photos is more than four, the change of calibration error is little. Weproposed that six photos are used for camera calibration. Calibration error affectedby the distance both the camera and the object is small, while detection precision ofcalibration reference point is high. We proposed that moderate distance is adoptedfor various view environments.(4)3D reconstruction is a key step of photogrammetry at traffic accident scene.With the location of the planar pattern both stand and flat, stereo vision is used toreconstruct 3D point .The experiment results is relatively reliable. The results withobject flat are better.