Monocular Vehicle 3D Bounding Box Estimation Using Homograhy and Geometry in Traffic Scene
Yiqiang Chen, Feng Liu, Pei Ke
Abstract
Video surveillance applications such as vehicle speed measurement and traffic condition monitoring are prevailing nowadays. Monocular 3D object detection by traffic surveillance cameras is one of the key means to achieve these functions. The methods in literature depend on camera calibration and require 3D annotation to learn a model. In this paper, we propose a novel vehicle 3D bounding box estimation method making use of the 3D-2D geometry consistency and homography transformation. Our method conducts vehicle 3D bounding box estimation using uncalibrated traffic cameras without requiring any 3D annotation and large computational resources. With quantitative and qualitative experiments, we show a comparable result to the state-of-that art methods and a faster execution time.