An improved point cloud denoising method in adverse weather conditions based on PP-LiteSeg network

Area of Science:

• Computer Vision
• Autonomous Driving Systems
• Sensor Data Processing

Background:

• LiDAR point cloud data (PCD) is vital for autonomous driving perception.
• Adverse weather degrades LiDAR PCD quality, impacting detection range and introducing noise.
• Existing denoising algorithms struggle with accuracy and speed in challenging conditions.

Purpose of the Study:

• To propose a real-time point cloud denoising network (SAFDN) for adverse weather conditions.
• To enhance feature extraction and segmentation accuracy for noisy LiDAR data.
• To improve the overall performance of autonomous driving perception systems.

Main Methods:

• Introduced the WeatherBlock module with dilated convolutions for enhanced receptive fields and multi-scale feature extraction.
• Developed the Series Attention Fusion Module (SAFM) using sequential channel and spatial attention mechanisms.
• Implemented weighted feature fusion to integrate low-level and high-level feature information.

Main Results:

• The SAFDN model demonstrated an 11.1% increase in denoising accuracy (MIOU) compared to the baseline PP-LiteSeg.
• Achieved a high inference speed of 205.06 FPS, enabling real-time processing.
• Significantly improved noise recognition accuracy and denoising capabilities in simulated adverse weather.

Conclusions:

• The proposed SAFDN effectively addresses the limitations of existing point cloud denoising methods in adverse weather.
• SAFDN offers a robust solution for enhancing LiDAR data quality, crucial for safe autonomous driving.
• The model's real-time performance makes it suitable for practical deployment in autonomous systems.