Vehicle Three-Dimensional Pose and Shape Estimation from Multiple Monocular Vision.

This paper proposes an accurate approach to estimate vehicles' 3D pose and shape from multi-view monocular images with a small overlap. This approach utilizes a state-of-the-art convolutional neural network (CNN) to detect vehicles' semantic keypoint in images and then introduces a Cross Projection Optimization (CPO) method to estimate the 3D pose accurately. During the iterative CPO process, it implements a new vehicle shape adjustment method named Hierarchical Wireframe Constraint (HWC). The approach is tested under both simulated and real-world scenes for performance verification. It's shown that this approach outperforms other existing monocular and stereo visual methods for vehicles' 3D pose and shape estimation. This approach provides new and robust solutions for accurate visual vehicle localization and it can be applied to the massive surveillance camera networks for intelligent transportation applications such as automatic driving assistance.