Abstract
Multi-modal sensors are the key to ensuring the robust and accurate operation of autonomous driving systems, where LiDAR and cameras are important on-board sensors. However, current fusion methods face challenges due to inconsistent multi-sensor data representations and the misalignment of dynamic scenes. Specifically, current fusion methods either explicitly correlate multi-sensor data features by calibrating parameters, ignoring the feature blurring problems caused by misalignment, or find correlated features between multi-sensor data through global attention, causing rapidly escalating computational costs. On this basis, we propose a transformer-based end-to-end multi-sensor fusion framework named the adaptive fusion transformer (AFTR). The proposed AFTR consists of the adaptive spatial cross-attention (ASCA) mechanism and the spatial temporal self-attention (STSA) mechanism. Specifically, ASCA adaptively associates and interacts with multi-sensor data features in 3D space through learnable local attention, alleviating the problem of the misalignment of geometric information and reducing computational costs, and STSA interacts with cross-temporal information using learnable offsets in deformable attention, mitigating displacements due to dynamic scenes. We show through numerous experiments that the AFTR obtains SOTA performance in the nuScenes 3D object detection task (74.9% NDS and 73.2% mAP) and demonstrates strong robustness to misalignment (only a 0.2% NDS drop with slight noise). At the same time, we demonstrate the effectiveness of the AFTR components through ablation studies. In summary, the proposed AFTR is an accurate, efficient, and robust multi-sensor data fusion framework.