A Survey of Deep Learning-Driven 3D Object Detection: Sensor Modalities, Technical Architectures, and Applications.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Robotics

Background:

• Deep learning has revolutionized 3D object detection, yet integrating diverse sensor data and architectures remains challenging.
• Existing surveys often focus on single sensor types or architectures, lacking a unified framework for synergistic analysis.

Purpose of the Study:

• To provide a comprehensive survey of deep learning-driven 3D object detection methods.
• To systematically analyze methods using a dual-axis framework: sensor modality and technical architecture.
• To explore applications and future research directions in 3D perception.

Main Methods:

• Classification of 3D object detection methods based on sensor modalities (RGB cameras, LiDAR, multimodal fusion).
• Analysis of technical architectures including convolutional networks, bird's-eye view (BEV) methods, and occupancy networks.
• Examination of cross-modal fusion paradigms at data, feature, and result levels.

Main Results:

• Evolution of LiDAR processing from voxel-based to pillar-based modeling.
• Advancements in spatiotemporal modeling for dynamic scene understanding using BEV and temporal fusion.
• Demonstration of three-level cross-modal fusion strategies for enhanced detection accuracy.

Conclusions:

• Synergistic innovation between sensor modalities and technical architectures is key to advancing 3D object detection.
• Future work should focus on depth perception, open-scene modeling, and lightweight deployment for broader applications.
• The reviewed methods show significant potential for autonomous driving and agricultural robotics.