AEFusion: An Attention-Based Ensemble Learning Approach for BEV Fusion Perception in Autonomous Modular Buses

Abstract

Autonomous modular buses (AMB) are considered a promising solution to the challenges in public transportation, as they can reduce commute times, enhance transfer convenience, and address supply-demand imbalances in transportation systems. Nonetheless, current research mainly focuses on operational aspects, whereas the high precision required for in-transit docking remains a critical challenge for implementation. The accuracy of current autonomous driving perception systems is often limited due to errors introduced by multi-sensor fusion methods. To address this issue, this paper introduces an attention-based ensemble learning fusion method (AEfusion) which includes a supervision module that utilizes the more accurate depth information from LiDAR to guide the generation of image depth information. Additionally, the fusion module incorporates two enhanced channel attention blocks and a spatial attention block to strengthen feature learning and integration. Experiments on both the nuScenes dataset and a self-collected dataset demonstrate that our method is suited for full-range docking perception in AMBs and is superior to the existing approaches.