Efficient BEV-based 3D detection with vehicle kinematic modeling

Engineering notes

Key topics: autonomous driving system, autonomous driving, bev, 3d detection, nuscenes, deployment, perception. See the paper for implementation details and experimental results.

Chinese explanation / 中文解读

中文解读待补充：本站会优先为端到端自动驾驶、BEV感知、3D目标检测、轨迹预测、路径规划、LiDAR感知等高价值论文补充中文说明。

Original abstract

Efficient and stable three-dimensional perception is essential for practical autonomous driving systems. Bird’s-Eye View (BEV)-based multi-camera 3D detection has become a dominant paradigm due to its unified spatial representation; however, existing approaches still face limitations in computational efficiency and temporal stability, hindering real-time deployment. This paper presents an efficient BEV-based 3D detection framework that improves inference efficiency, orientation reliability, and temporal consistency through lightweight network design, motion-aware supervision, and vehicle kinematic modeling. At the network level, a lightweight convolution module, termed DSConv-SE², is introduced to enhance depth-wise separable convolution via dual-path squeeze and excitation channel modulation applied before and after convolution. This design strengthens inter-channel dependencies while maintaining low computational cost, resulting in more stable feature representations and reduced floating-point operations, particularly in high-level backbone stages and feature pyramid networks. To improve orientation robustness, a velocity-aware orientation re-weighting strategy is proposed, which adjusts yaw supervision according to object motion state. By suppressing ambiguous supervision for stationary vehicles and emphasizing heading consistency for moving ones, the method stabilizes orientation regression without additional inference overhead. Furthermore, a category-adaptive single-track vehicle kinematic model combined with an extended Kalman filter is incorporated to enhance temporal consistency. By integrating detection outputs with motion priors such as trajectory curvature and steering dynamics, the framework produces smoother and more physically plausible pose trajectories with reduced temporal jitter. Experiments on the nuScenes dataset demonstrate that the proposed method reduces computational cost with minimal accuracy degradation while improving orientation accuracy and temporal stability, making it suitable for real-time and resource-constrained autonomous driving applications.

Links and sources

Comments