CenterLoc3D: monocular 3D vehicle localization network for roadside surveillance cameras

Abstract Monocular 3D vehicle localization is an important task for behaviour analysis, traffic flow parameter estimation and autonomous driving in Intelligent Transportation System (ITS) Cooperative Vehicle Infrastructure (CVIS), which usually achieved by monocular detection. However, cameras cannot obtain depth information directly due to the inherent imaging mechanism, resulting more challenging tasks. Currently, most of detection methods still rely on 2D detectors additional geometric constraint modules recover information, reduces efficiency. At same time, research based datasets onboard scenes, instead roadside perspective, limited large-scale perception. Therefore, we focus without scenes. We propose a network CenterLoc3D cameras, predicts centroid eight vertexes image space, dimension bounding boxes detectors. To improve precision localization, multi-scale weighted-fusion module loss with spatial constraints embedded CenterLoc3D. Firstly, transformation matrix between space world solved camera calibration. Secondly, type, centroid, vertexes, are obtained Finally, can be calibration localization. best our knowledge, this first application cameras. Hence, also benchmark including dataset (SVLD-3D), annotation tool (LabelImg-3D), evaluation metrics. Through experimental validation, proposed method achieves high accuracy $$A{P_{3D}}$$ A P 3 D 51.30%, average 98%, 85% real-time performance FPS 41.18.