3D vision object detection for autonomous driving in fog using LiDaR

Abstract

Connected and Autonomous Vehicles (CAVs) are transforming transportation. The paper describes a new method of fog simulation applied to LiDAR data for self-driving cars with a focus on enhancing 3D object detection in low visibility conditions. As opposed to the previously used methods, synthetic fog augmentation is combined with deep learning models and it is proven that the proposed method is superior to the previous methods when it comes to object detection accuracy in various fog levels. Another challenge that has been discussed in the study to ensure the reliability of autonomous navigation is the question of how the fog and the LiDAR point cloud should be modeled which eventually helps in improving the decision-making safety and operation. Fog can drastically reduce visibility and safety, making it crucial to test LiDAR-based perception algorithms for CAVs under such conditions. These simulations aim to ensure CAVs can navigate safely and efficiently through fog. However, challenges like sensor calibration and data integration need to be addressed. Despite these hurdles, the research foresees a future where CAVs, equipped with advanced LiDAR-based perception algorithms and fog-handling capabilities, enhance safety and efficiency in transportation. Notably, using synthetic fog augmentation improved detection by 5.27% for cars and 8.11% for cyclists. Furthermore, the study showcases improvements of 4.76%, 2.92%, and 3% in Mean Average Precision (mAP) across the distinct object categories of easy, moderate, and hard difficulty levels, respectively.