Grid-Centric Traffic Scenario Perception for Autonomous Driving: A Comprehensive Review.

Area of Science:

• Robotics and Autonomous Systems
• Computer Vision
• Artificial Intelligence

Background:

• Grid-centric perception is vital for mobile robot navigation but less common than object-centric methods due to complexity.
• Autonomous vehicles require accurate perception in dynamic, large-scale traffic, posing challenges for traditional approaches.
• Advancements in deep learning and hardware are driving innovation in grid-centric perception.

Purpose of the Study:

• To provide a comprehensive, hierarchically structured review of grid-centric perception for autonomous vehicles.
• To bridge the knowledge gap in this rapidly evolving research area.
• To consolidate current understanding and identify future research directions.

Main Methods:

• Systematic review of existing literature on grid-centric perception techniques.
• Hierarchical organization of knowledge from 2D Bird's-Eye View (BEV) grids to 3D occupancy and 4D occupancy forecasting.
• Summarization of label-efficient learning and the role of grid-centric perception in driving systems.

Main Results:

• Grid-centric perception offers advantages like fine-grained environmental representation, robustness to occlusion, and improved ground estimation.
• Occupancy networks are expanding to 4D scene perception and prediction, incorporating topics like generative AI and world models.
• The review covers techniques from 2D BEV to 3D and 4D occupancy forecasting, including label-efficient learning.

Conclusions:

• Grid-centric perception, with its geometry-first paradigm, is more robust for open-world driving scenarios with unknown obstacles.
• Emerging trends include 4D occupancy forecasting, generative AI, and world models, significantly enhancing autonomous driving capabilities.
• This review provides a foundational understanding and outlook for future research in grid-centric perception for autonomous vehicles.