Lifting 2D Object Detection to 3D: Geometric Approach in Bird-Eye-View

3D object detection is fundamental for computer vision applications. Although motion is essentially linked with the orientation of the object, the trajectory of the object is not fully utilized in modern 3D object detectors. This paper proposes a method for object 3D bounding box estimation based on 2D tracking and Inverse Perspective Mapping. Unlike other solutions utilizing deep learning techniques for orientation estimation, the proposed method uses geometric constraints and object movements. Which causes the execution speed and errors of 3D localization to remain comparable to a pure 2D tracker.