Simulation and Performance Analysis of 3D Object Detection Algorithm using Deep Learning for Computer Vision Applications

Abstract

3D object detection is becoming one of the main areas of research centered around Autonomous Vehicles. It involves real-time detection and tracking of 3D objects in a scenario. With the help of autonomous driving vehicles, one can ensure safety of human life and decrease the number of fatalities on road due to human error. Identified the problem with 3D object detection being that there is still a gap between 3D and 2D architecture. It is difficult to extract 3D object bounding boxes from monocular images because 2D images has lack of depth information, which is necessary for 3D object detection. In order to overcome this, a LiDAR approach is being used, which is based on the use of convolutional neural networks (CNNs) to get depth information. With one more dimension, the estimated depth map adds more information for 3D detection. Proposed work evaluated on standard benchmark datasets like KITTI dataset. 3D object detection systems are designed to provide 3D-oriented bounding boxes for 2D objects in 2D images. 8-corner, 3D centers with offsets, and 4-corner-2-height representations are used to parameterize 3D cuboids. Firstly, a 2D bounding box is created around the object estimated from the image. Then, the 3D LiDAR cloud points are superimposed on the image as an intermediate step to aid in 3D estimation. Using the cloud points, the depth map is formed and a 3D bounding box is created in x, y, and z plane. The proposed object detection model obtained an average accuracy of 98.1% on KITTI and custom dataset.