Towards Real-Time 3D Object Detection Through Inverse Perspective Mapping

3D object detection in road scenes is crucial for computer vision applications. While traditional 2D object detection has seen remarkable success, 3D object detection is still an active area of research. This paper presents an efficient method for monocular 3D object detection using inverse perspective mapping. The proposed method differs from other approaches that rely on deep learning techniques to estimate an object’s orientation. Instead, it uses geometric constraints to lift 2D detections to 3D. The method utilizes instance segmentation and derived object dimensions to achieve 3D localization by employing inverse perspective mapping. The method successfully combines these elements to determine the object’s position in 3D space accurately. By doing so, it achieves execution speeds and localization errors comparable to pure 2D methods, while also offering the advantages of 3D detection. This work defines real world scenarios such as for roadside surveillance cameras and onboard cameras, for which high accuracy and speed of 3D localization can be guaranteed by the proposed algorithm.