Deep Learning for LiDAR Point Clouds in Autonomous Driving: A Review.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Robotics

Background:

• Autonomous driving relies heavily on 3-D LiDAR data for environmental perception.
• Processing uneven, unstructured, and noisy 3-D point clouds presents significant challenges for automated systems.
• Deep learning (DL) has shown promise in extracting discriminative features from 3-D LiDAR data.

Purpose of the Study:

• To provide a systematic review of deep learning architectures applied to 3-D LiDAR point clouds for autonomous driving tasks.
• To bridge the gap in existing literature by offering a comprehensive survey on this specific topic.
• To consolidate recent advancements and identify future research directions.

Main Methods:

• Systematic literature review of deep learning architectures for 3-D LiDAR point cloud processing.
• Categorization of DL applications into segmentation, detection, and classification tasks relevant to autonomous driving.
• Summary and analysis of over 140 key contributions from the past five years.

Main Results:

• Detailed overview of milestone 3-D deep learning architectures.
• Comprehensive summary of DL applications in 3-D semantic segmentation, object detection, and classification.
• Inclusion of relevant datasets, evaluation metrics, and state-of-the-art performance benchmarks.

Conclusions:

• Deep learning has significantly advanced autonomous driving perception using 3-D LiDAR data.
• Despite progress, challenges remain in automated processing of complex point clouds.
• Future research should focus on addressing these challenges and further enhancing DL model performance.