3D Pose Estimation of Vehicles Using Stereo Camera

Excerpt

For advanced driver assistance systems, the 3D poses and motion states of oncoming and intersecting vehicles represent important information. This work describes methods for 3D vehicle pose estimation based on a motion-attributed 3D point cloud generated. First, stereo and optical flow information is computed for the investigated scene. A four-dimensional clustering approach separates the static from the moving objects in the scene. The iterative closest point algorithm (ICP) estimates the vehicle pose using a cuboid as a weak vehicle model. Classical ICP optimization is based on the Euclidean distance metric. Its computational efficiency can be significantly increased by applying a quaternion-based optimization scheme. In vehicle-based small-baseline stereo systems, it is favorable to use a polar distance metric which especially takes into account the error distribution of the stereo measurement process. To derive the pose parameters and simultaneously their temporal derivatives, i.e., the motion state of the vehicle, from two subsequent stereo image pairs in order to avoid a temporal filtering stage, a model-based scene flow method is applied. Here, 3D points on the object surface are reprojected into the stereo images and similarities between image windows extracted around the reprojected points are maximized. An experimental evaluation is performed on seven different real-world sequences, where different stereo algorithms, baseline distances, distance metrics, and optimization algorithms are examined. The results show that the proposed polar distance metric yields a higher accuracy especially for the estimation of the yaw angle of a vehicle than the common Euclidean distance metric, especially when using pixel-accurate stereo points. The model-based scene flow approach yields a refined 3D pose estimation along with the instantaneous motion state of the vehicle. …