Highly Efficient MetaFormer-Based End-to-End Autonomous Driving Model With Token-Pruning

Engineering notes

These results suggest that the proposed approach effectively resolves the trade-off between real-time performance and recognition accuracy, demonstrating its superiority in autonomous racing.

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

In autonomous racing, high-speed and accurate environmental perception is crucial. This study focuses on the Vision Transformer (ViT) based MetaFormer and aims to achieve real-time performance and high recognition accuracy by applying token-pruning method. The proposed approach applies token-pruning to specific stages within the MetaFormer, dynamically reducing tokens to decrease computational complexity. In the image recognition task using the ImageNet dataset, the proposed method achieved up to a 7.5% reduction in computational cost compared to the baseline MetaFormer. Furthermore, in the autonomous driving task using a simulator, the proposed method attained a reduction of 0.33 seconds in the fastest lap-time and 0.93 seconds in the total time over a maximum of three laps. These results suggest that the proposed approach effectively resolves the trade-off between real-time performance and recognition accuracy, demonstrating its superiority in autonomous racing.

Links and sources

• Official / arXiv page

Comments