LST-BEV: Generating a Long-Term Spatial-Temporal Bird's-Eye-View Feature for Multi-View 3D Object Detection.

Area of Science:

• Computer Vision
• Autonomous Driving Systems
• Machine Learning

Background:

• Traditional 3D object detection often relies on LiDAR, but multi-camera visual perception offers a cost-effective alternative.
• Existing visual perception methods face challenges in capturing long-range spatial-temporal dependencies and integrating cross-task information.

Purpose of the Study:

• To develop a novel multi-view 3D object detection framework, LST-BEV, enhancing performance for autonomous driving applications.
• To address limitations in current attention mechanisms for capturing long-range dependencies and cross-task information.

Main Methods:

• Proposed a Long-Range Cross-Task Detection Head (LRCH) to capture long-range dependencies and integrate cross-task information.
• Introduced a Long-Term Temporal Perception Module (LTPM) combining Mamba and linear attention for efficient temporal feature extraction.

Main Results:

• The LST-BEV framework demonstrated significant performance improvements on the nuScenes dataset.
• Achieved a 2.1% increase in mean Average Precision (mAP) and a 2.7% increase in nuScenes Detection Score (NDS) compared to the SA-BEVPool baseline.

Conclusions:

• The proposed LST-BEV framework effectively enhances multi-view 3D object detection for autonomous driving.
• The novel LRCH and LTPM modules contribute to overcoming limitations in existing methods, leading to superior detection performance.