TY - GEN
T1 - A Lightweight YOLOv8-Based Camera Backbone for Efficient BEVFusion
AU - Zhu, Yifan
AU - Jin, Hui
N1 - Publisher Copyright:
© 2026 Copyright held by the owner/author(s).
PY - 2026/8/19
Y1 - 2026/8/19
N2 - To address the high computational cost and limited real-time performance of multimodal 3D perception systems, this paper explores camera backbone lightweight design for efficient multi-view feature extraction. A lightweight YOLOv8-based backbone is first constructed through structural optimization, and then integrated into the BEVFusion framework by replacing the original Swin Transformer-based camera encoder without modifying the overall fusion pipeline. On the nuScenes benchmark, the proposed backbone demonstrates competitive 3D detection accuracy alongside improved inference efficiency, despite requiring considerably fewer parameters and lower computational cost. This confirms that the method achieves a favorable balance between performance and efficiency, making it particularly suitable for real-time autonomous driving scenarios where hardware resources are constrained.
AB - To address the high computational cost and limited real-time performance of multimodal 3D perception systems, this paper explores camera backbone lightweight design for efficient multi-view feature extraction. A lightweight YOLOv8-based backbone is first constructed through structural optimization, and then integrated into the BEVFusion framework by replacing the original Swin Transformer-based camera encoder without modifying the overall fusion pipeline. On the nuScenes benchmark, the proposed backbone demonstrates competitive 3D detection accuracy alongside improved inference efficiency, despite requiring considerably fewer parameters and lower computational cost. This confirms that the method achieves a favorable balance between performance and efficiency, making it particularly suitable for real-time autonomous driving scenarios where hardware resources are constrained.
KW - 3D Object Detection
KW - BEVFusion
KW - Lightweight Backbone
KW - Multimodal Perception
KW - YOLOv8
UR - https://www.scopus.com/pages/publications/105048677544
U2 - 10.1145/3810280.3810313
DO - 10.1145/3810280.3810313
M3 - Conference contribution
AN - SCOPUS:105048677544
T3 - Proceedings of 2026 2nd International Conference on Artificial Intelligence and Computational Intelligence, AICI 2026
SP - 228
EP - 235
BT - Proceedings of 2026 2nd International Conference on Artificial Intelligence and Computational Intelligence, AICI 2026
PB - Association for Computing Machinery, Inc
T2 - 2026 2nd International Conference on Artificial Intelligence and Computational Intelligence, AICI 2026
Y2 - 6 February 2026 through 8 February 2026
ER -