Lidar-Based Car Detection and Position Estimation Using Convolutional Neural Networks for Autonomous Driving

Abstract

Despite the many applications of computer vision in robotics, automation, and self-driving cars, object detection remains a significant difficulty. Recent developments in object recognition and image classification, in particular convolutional neural networks (CNNs), have significantly aided in the development of neural networks that enable practical applications in autonomous driving. This study uses a modified CNN architecture designed for accurate 3-D item identification to examine vehicle recognition and spatial analysis utilizing LiDAR data. The network offers robust localization even in difficult environments with occlusions using point-based bounding box predictions. Using a RANSAC-based refinement technique, the suggested system achieved a classification accuracy of 96.71% with bounding box precision significantly enhanced after 246 epochs of extensive evaluation on the KITTI dataset. These results validate the model's strong generalization ability in real-world scenarios with mirrored, noisy, and variable-oriented data. The findings validate the growing significance of CNNs in addressing autonomous navigation challenges. By contributing to the development of safer and more reliable transportation networks, our work ultimately moves the road toward completely autonomous vehicles.