Lidar waveforms are worth 40x128x33 words
2 days ago
- #Autonomous Driving
- #Lidar
- #Transformer Architecture
- Lidar is crucial for autonomous driving, providing high-resolution 3D scans for scene understanding.
- Conventional automotive lidar DSPs process waveforms individually, missing valuable context from neighboring waveforms.
- Lidar point clouds are prone to artifacts from low SNR regions, reflective objects, and environmental conditions like fog.
- A learned DSP using a transformer architecture processes full waveforms, leveraging adjacent waveforms for high-fidelity multi-echo point clouds.
- The method improves Chamfer distance by 32cm and 20cm compared to conventional peak finding and transient imaging approaches.
- Maximum range improvements include up to 17m in fog and 14m in nominal real-world conditions.