Free Space Detection Using Camera-LiDAR Fusion in a Bird's Eye View Plane

Area of Science:

• Computer Vision
• Robotics
• Autonomous Systems

Background:

• Vision-based road segmentation faces challenges with sensor noise, object deformation, and low-resolution long-distance imagery.
• Distinguishing roads from other objects in complex visual data remains a significant hurdle for autonomous systems.

Purpose of the Study:

• To develop a robust road detection system by effectively fusing camera and Light Detection and Ranging (LiDAR) data.
• To overcome limitations of vision-only approaches by leveraging LiDAR's geometric and intensity information.

Main Methods:

• A novel convolutional neural network (CNN) architecture processes data transformed into a bird's eye view (BEV) space.
• The network incorporates dual pathways to address calibration errors between image and point cloud data.
• Sequential modules with varying dilated convolution rates enable efficient processing of diverse data scales.

Main Results:

• The proposed method achieves robust road detection by integrating camera and LiDAR sensor information.
• Empirical evaluations on the KITTI road detection benchmark demonstrate competitive performance.
• The system achieves real-time processing speeds, ranking 22nd on the KITTI leaderboard.

Conclusions:

• Fusing camera and LiDAR data with a BEV-transformed CNN significantly improves road detection robustness and accuracy.
• The developed architecture effectively handles calibration issues and processes multi-scale data efficiently.
• The real-time performance of the system supports its application in practical autonomous driving scenarios.