基于精细模型的视频车辆三维姿态估计

3D POSE ESTIMATION OF VIDEO VEHICLE BASED ON FINE MODEL

由于实际监控场景中包含大量车辆遮挡、复杂光照、运动模糊等干扰因素,导致监控视频中车辆的三维姿态估计精度低下。针对该问题,提出一种基于精细三维模型的车辆姿态估计方法。利用模型车辆轮廓与图像中真实车辆轮廓的匹配误差构建能量函数,并采用高斯牛顿算法优化姿态参数,求解能量函数最小化问题得到最终结果。轮廓匹配时,采用改进的随机抽样一致性(RANSAC)算法对曲线模型轮廓用分段直线拟合,再与真实车辆轮廓匹配。实验结果表明,该算法的准确性相较于目前方法有明显提升,并在复杂环境中有很好的鲁棒性。

Due to the vehicle occlusion,complex illumination,motion blur,and other factors,the accuracy of the three-dimensional pose estimation of the road vehicle is worsened. Aiming at this issue,a vehicle poses estimation algorithm based on fine 3 D model was proposed. The energy function was constructed by using the matching error between the model vehicle contour and the image vehicle contour. Gauss-Newton algorithm was used to optimize the pose parameters to minimize the energy function,and got the final convergence poses. For contour matching,the contour of the model vehicle was fitted with some straight line using an improved RANSAC algorithm and then matched with the image vehicle contour. Experimental results show that the accuracy of the algorithm is obviously improved compared with the present method,and it has good robustness in complex environment.