3D Object Detection with SLS-Fusion Network in Foggy Weather Conditions.

Area of Science:

• Computer Vision
• Robotics
• Autonomous Systems

Background:

• Sensors like cameras and LiDAR are vital for self-driving car environmental awareness.
• Extreme weather, particularly fog, distorts sensor data, posing safety risks.
• Existing 3D object detection methods struggle in adverse weather conditions.

Purpose of the Study:

• To analyze fog's effect on 3D object detection in driving scenes.
• To propose and validate methods for improving detection performance in foggy conditions.
• To develop a synthetic dataset for simulating fog effects.

Main Methods:

• Fog synthesis applied to the public KITTI dataset to create the Multifog KITTI dataset (images and point clouds).
• Evaluation of the Spare LiDAR Stereo Fusion Network (SLS-Fusion) 3D object detector on original and augmented datasets.
• A specific training strategy using both original and augmented data to enhance robustness.

Main Results:

• Performance in 3D object detection for Moderate objects decreased by 42.67% in foggy conditions without improvements.
• The proposed training strategy significantly improved performance by 26.72% in foggy conditions.
• The algorithm maintained good performance on the original dataset with only an 8.23% drop.

Conclusions:

• Fog severely degrades 3D object detection capabilities in autonomous driving systems.
• Additional training with a synthetic fog dataset substantially enhances the robustness of 3D object detection algorithms.
• The developed methods offer a viable solution for improving self-driving car safety in foggy weather.