A Review of Multi-Sensor Fusion in Autonomous Driving.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Robotics

Background:

• Multi-modal sensor fusion is critical for autonomous driving perception.
• Deep learning methods are increasingly used to integrate data from cameras, LiDAR, and radar.

Purpose of the Study:

• To provide a structured overview of recent deep learning-based sensor fusion techniques.
• To analyze architectural trends, learning strategies, and applications in autonomous driving.

Main Methods:

• Categorization of fusion methods by architectural paradigms (e.g., BEV-centric, cross-modal attention).
• Review of learning strategies and task adaptations for perception tasks.
• Analysis of dominant trends like unified Bird's-Eye View (BEV) representation and token-level alignment.

Main Results:

• Identified two key architectural trends: unified BEV representation and token-level cross-modal alignment.
• Reviewed diverse applications including object detection, semantic segmentation, behavior prediction, and planning.
• Highlighted challenges hindering real-world deployment: spatio-temporal misalignment, domain shifts, and interpretability.

Conclusions:

• Future research directions include diffusion models, Mamba-style architectures, and large vision-language models for scalable and trustworthy perception.
• Extensive comparisons and benchmark analyses are provided to guide future research.
• Addressing current challenges is key for advancing robust autonomous driving systems.