Real-Time 3D Bounding Box Estimation with RCNN-Resnet101 and Adaptive Projection Matrices

Abstract

In recent years, there has been a surge in demand for precise object detection and 3D bounding box estimation across various applications in computer vision, ranging from autonomous driving to robotics and augmented reality. This paper introduces a method specifically designed for real-time 3D bounding box estimation using a monocular camera. In contrast to conventional approaches that directly estimate 3D bounding boxes, our method begins by calculating accurate 2D bounding boxes. Subsequently, it employs a combination of mathematical modelling to predict 3D bounding boxes. The primary goal of this research is to deliver a practical and efficient solution for achieving accurate 3D bounding box estimation, capitalizing on the widespread availability of monocular camera hardware. Within our study, we independently generate the Projection Matrix, ensuring the robustness and applicability of our method across diverse datasets. This underscores the effectiveness of our approach in accurately estimating 3D object properties. Notably, our methodology demonstrates satisfactory results on both the Kitti and Cityscapes datasets, further emphasizing its proficiency in accurately estimating 3D object properties.