Efficient Stereo Depth Estimation for Pseudo-LiDAR: A Self-Supervised Approach Based on Multi-Input ResNet Encoder.

Area of Science:

• Computer Vision
• Robotics
• Autonomous Systems

Background:

• Autonomous delivery vehicles rely on 3D LiDAR for precise perception and localization.
• Existing camera-based methods struggle with real-time performance compared to LiDAR.

Purpose of the Study:

• To develop a real-time pseudo point cloud generation strategy using image sensors.
• To enable autonomous vehicles to achieve LiDAR-like perception and localization capabilities from camera data.

Main Methods:

• Utilized stereo imagery with advanced depth estimators for accurate depth map generation.
• Developed a novel approach to create pseudo point clouds from estimated depth maps.
• Focused on achieving real-time processing speeds.

Main Results:

• The proposed method generates pseudo point clouds significantly faster than existing approaches.
• Achieved superior performance in depth map generation on the KITTI dataset.
• Demonstrated effectiveness and real-time capabilities on the KITTI stereo benchmark.

Conclusions:

• The camera-based pseudo point cloud generation method offers a viable, high-performance alternative to LiDAR for autonomous vehicle perception.
• This approach enhances the feasibility of real-time localization and navigation for autonomous delivery vehicles.