A 3D object detection strategy based on YOLOv8 combined with CaDDN for vehicle images

With the rapid development of deep learning and computer vision technologies, 3D object detection has become more and more widely used in fields such as autonomous driving, robotics and augmented reality. However, 3D object detection from monocular images is still a challenging problem. To solve this problem, this paper proposes a new method combining YOLOv8 and CaDDN (abbreviated as YO-Ca). First, we use YOLOv8 to perform 2D object category detection on monocular images, providing a 2D bounding box for each detected vehicle. Then, we employ the linear increasing discretization (LID) approach in the CaDDN method to predict the depth distribution of each pixel. In this way, we can accurately localize the image information in 3D space. Then, we convert the 2D bounding box into a 3D bounding box in the bird's-eye view, and refine and verify the 3D bounding box in the bird's-eye view. The experimental results show that our proposed YO-Ca method achieves better performance on the KITTI dataset, especially in terms of 3D intersection and merger ratio (IoU=0.7).