A Deep Learning Framework for Accurate Vehicle Yaw Angle Estimation from a Monocular Camera Based on Part Arrangement

An accurate object pose is essential to assess its state and predict its movements. In recent years, scholars have often predicted object poses by matching an image with a virtual 3D model or by regressing the six-degree-of-freedom pose of the target directly from the pixel data via deep learning methods. However, these approaches may ignore a fact that was proposed in the early days of computer vision research, i.e., that object parts are strongly represented in the object pose. In this study, we propose a novel and lightweight deep learning framework, YAEN (yaw angle estimation network), for accurate object yaw angle prediction from a monocular camera based on the arrangement of parts. YAEN uses an encoding−decoding structure for vehicle yaw angle prediction. The vehicle part arrangement information is extracted by the part-encoding network, and the yaw angle is extracted from vehicle part arrangement information by the yaw angle decoding network. Because vehicle part information is refined by the encoder, the decoding network structure is lightweight; the YAEN model has low hardware requirements and can reach a detection speed of 97FPS on a 2070s graphics cards. To improve the performance of our model, we used asymmetric convolution and SSE (sum of squared errors) loss functions of adding the sign. To verify the effectiveness of this model, we constructed an accurate yaw angle dataset under real-world conditions with two vehicles equipped with high-precision positioning devices. Experimental results prove that our method can achieve satisfactory prediction performance in scenarios in which vehicles do not obscure each other, with an average prediction error of less than 3.1° and an accuracy of 96.45% for prediction errors of less than 10° in real driving scenarios.