3D Pose Estimation Of Vehicle Image Based On Fine Models

In order to ensure urban security,there have been more and more surveillance cameras,resulting in a huge amount of surveillance video data.It takes a lot of resources to analyze and process these massive data.The current computer vision method can be used to automate the surveillance video and obtain effective information from it.In the surveillance video,the road vehicle is an important detection object,and accurately estimating the 3D posture parameters of the vehicle relative to the camera in the video is of great significance for analyzing and understanding the video content.Estimating the 3D pose of an object in image is a classical issue in computer vision.Due to the loss of depth information in the captured image,a large amount of detailed information disappears,making the problem very difficult.We take the road vehicle as the research object,and intends to estimate the pose of the vehicle in the surveillance video.Due to the vehicle occlusion,complex illumination,motion blur,and other factors,the accuracy of the 3D estimation of the road vehicle is worsened.In order to solve this problem,a vehicle pose estimation algorithm based on fine 3D model is proposed.The fine 3D models are used as prior information.We use the edge contour feature of the vehicle to establish the correspondence between the 2D image and 3D model.The energy function is constructed by using the matching error between the model vehicle contour and the image vehicle contour.Gauss-Newton algorithm is used to optimize the pose parameters to minimize the energy function,and get the final convergence pose.For contour matching,in order to avoid curve contour matching,the contour of the model vehicle is fitted with some straight line using an improved RANSAC algorithm and then matched with the image vehicle contour.In the optimization process,the optimization techniques such as weighting and multi-scale optimization are used to make the results more accurate and robust.Finally,we use the estimation of distribution algorithms to directly solve the six-dimensional pose parametersWe carried out experiments in a simulated monitoring environment.The experimental environment is controllable and can simulate various complex monitoring environments.Experimental results show that the proposed method compared to the current method can significantly improve the accuracy and robustness.