基于一维直接线性变换的视频中车辆运动状态重建

Reconstruction of Vehicle Movement in Video Sequences Based on One-dimensional Direct Linear Transformation

为克服传统车辆运动状态重建方法不能全面反映视频图像中车辆运动状态,且使用条件受限较大的问题,基于近景摄影测量中的直接线性变换原理,结合车身外廓特征信息,提出一种完整重建视频中车辆运动状态的有效方法。该方法中的特征标定信息全部取自目标车辆的外廓特征,不受路面和环境标定条件影响,扩大了使用范围;标定区域覆盖车辆在视频中的整个运动过程,最大限度地保证了车辆行驶轨迹的空间完整性;方法中每相邻2帧之间车辆行驶距离、行驶速度及加速度的解算均独立,避免产生累计误差。最后,使用该方法分别对车辆处于低速、中高速或减速3种运动状态下,摄像方向与车辆行驶方向呈90°或30°夹角的6种组合试验中车辆的相关运动状态参数进行解算,并与试验中采集的实际运动状态参数进行分析对比。研究结果表明:当车辆分别处于低速、中高速或减速3种运动状态时,在90°摄像视角下,计算所得车速值与记录值误差在1.5%以内,行驶距离值误差在3%以内,加速度值误差在7%以内;在30°摄像视角下,计算所得车速值误差在4%以内,行驶距离值误差在5%以内,加速度值误差在9%以内;该方法计算的视频中车辆的车速和行驶距离精度较高,加速度精度满足相关行业应用要求,证明该方法用于重建视频中车辆的运动状态有效、可行。

To overcome the problem that previous calculation methods could not reconstruct vehicle movement comprehensively and their application conditions were significantly limited, a new calculation method based on the one-dimensional Direct Linear Transformation （DLT）, combined with the contour structure of the vehicle, was proposed to calculate the vehicle movement state in video sequences. The calibration information in the method was obtained from the contour structure of the target vehicle, which was not influenced by the calibration conditions of the road and environment, and this allowed expansion of its application. Meanwhile, the entire movement process of the vehicle in the video was fully encompassed by the calibration areas;therefore, the spatial wholeness of vehicle＇s movement track reconstruction was ensured. The calculations of the distance, speed, and acceleration were mutually independent, which eliminated cumulative errors. Finally, the confirmatory experiment of the vehicle driving in low speed, medium-high speed, and deceleration phase, as well as camera perspective directions of 90° and 30°, were conducted. The vehicle movements were calculated by this method, and then analyzed and compared with the real movement state recorded in the test. The results show that, in the camera perspective directions of 90° and 30°, respectively, regardless of the driving state, the errors between the calculated speed and the recorded speed were less than 1.5% and 4%, errors of distances were less than 3% and 5%, and errors of accelerations were less than 7% and 9%. This proves that the precision accuracies of both the speed and distance calculated by this method are improved greatly, and the accuracies of the calculated acceleration also satisfy the requirements. The results also demonstrate that the method is effective and practical in reconstructing vehicle movement in video sequences.