Fast-BEV: A Fast and Strong Bird's-Eye View Perception Baseline

Area of Science:

• Computer Vision
• Autonomous Driving Systems
• Machine Learning

Background:

• Bird's-Eye View (BEV) representation is crucial for autonomous vehicle (AV) perception.
• Existing BEV solutions often require significant computational resources or exhibit limited performance.
• There is a need for efficient and high-performing BEV perception systems deployable on-vehicle.

Purpose of the Study:

• To propose Fast-BEV, a novel framework for faster and deployment-friendly BEV perception on automotive chips.
• To demonstrate that powerful BEV representation can be achieved without complex transformers or depth estimation.
• To achieve a balance between high performance and fast inference speed for on-vehicle applications.

Main Methods:

• Developed a lightweight, deployment-friendly view transformation for efficient 2D-to-3D voxel feature transfer.
• Utilized a multi-scale image encoder to capture diverse feature information.
• Introduced an efficient BEV encoder optimized for on-vehicle inference speed.
• Implemented a robust data augmentation strategy for both image and BEV spaces.
• Incorporated a multi-frame feature fusion mechanism to leverage temporal information.

Main Results:

• The Fast-BEV framework achieves significantly faster inference speeds compared to existing methods (e.g., 52.6 FPS for R50 model).
• Maintained competitive performance, with the R50 model achieving 47.3% NDS on the nuScenes validation set.
• The largest model (R101@900×1600) reached 53.5% NDS on nuScenes, demonstrating strong accuracy.
• Developed a benchmark evaluating accuracy and efficiency on popular on-vehicle chips.

Conclusions:

• Fast-BEV provides a high-performance, fast, and deployment-friendly solution for autonomous driving perception.
• The framework effectively balances computational efficiency with robust perception capabilities.
• Empirically validated the efficacy of simplified BEV representation without transformers or explicit depth estimation.