3D Vehicle Information Recognition Algorithm of Monocular Camera Based onSelf-Calibration in Traffic Scene

Abstract

<p indent="0mm">Obtaining 3D information of vehicles as the basis for accurate classification of vehicles has become an increasingly important research direction. However, most of the current traffic monitoring cameras are monocular cameras, which cannot directly obtain 3D information of vehicles like pose and size due to perspective factors. According to the above problem, this paper proposes a 3D vehicle information recognition algorithm of monocular camera based on self-calibration in traffic scene. Firstly, this paper builds up a monocular camera model and a stable single vanishing point calibration model according to the typical traffic scene, and completes camera calibration. Then it uses the YOLO deep learning convolution neural network for 2D vehicle detection. Based on this, it puts forward a diagonal and vanishing point constrained non-linear optimization algorithm, combining with the calibration information to complete 3D vehicle information recognition and the best 3D vehicle detection. Finally, the experiment was carried out on the public dataset called BrnoCompSpeed and in highway traffic scenes, and the results show that the algorithm can effectively complete 3D vehicle information recognition in various traffic scenarios with an average recognition accuracy of more than 90%.