Recent research in autonomous driving perception is shifting from expensive LiDAR-dependent architectures to practical, cost-effective Bird’s-Eye-View (BEV) pipelines that satisfy real-time constraints. However, existing Transformer-based BEV transf...
Recent research in autonomous driving perception is shifting from expensive LiDAR-dependent architectures to practical, cost-effective Bird’s-Eye-View (BEV) pipelines that satisfy real-time constraints. However, existing Transformer-based BEV transformation techniques suffer from high computational complexity due to multiple decoder layers and complex attention structures, making real-time inference on encoding systems challenging. This paper proposes a lightweight BEV perception pipeline using only multi-view cameras, aiming to balance accuracy and real-time performance through a novel altitude-aware position encoding fusion technique.
The core of the proposed WAFormer-based View Transformer is to significantly reduce computational costs by vertically pooling multi-view image features into 1D feature maps. To recover information lost during vertical compression, a Refine Transformer utilizing Self-Attention and Cross-Attention—using original 2D features as Keys/Values—is employed. Furthermore, the model constructs an Altitude PE based on ground-relative altitude distribution, which is insensitive to distance, alongside standard depth-based embeddings. These embeddings are aggregated into 1D forms via HeightNet, fused through bidirectional Cross-Attention and LayerScale, and combined with multi-view position encoding to generate the final width-directional encoding. This enables WAFormer to generate lightweight 1D tokens that preserve both depth and altitude information while minimizing computation.
The overall network consists of a ResNet-101 backbone, the WAFormer transformer, a BEV encoder, and an FCOS-based detection head. Temporal information is incorporated by aligning past BEV features using ego-motion matrices to form 4D spatiotemporal features. The detection head treats each location on the feature map as a candidate point, simultaneously predicting class probability, centerness, categorical distributions of depth and height, and 3D bounding box size.
The proposed method was evaluated on the nuScenes dataset across various scenarios. In diverse environments the model showed consistent mAP and NDS improvements over BEVDet4D. Notably, it maintained stable performance improvements even in occlusion scenarios. Overall, the proposed model achieved higher accuracy and similar inference speeds compared to the baseline, confirming its capability for real-time inference unlike the high-latency BEVFormer.
This paper proposes a lightweight WAFormer-based BEV transformation structure incorporating vertical compression and depth-altitude fusion, experimentally verifying stable 3D object detection using only multi-view cameras. By achieving a balance between accuracy and real-time performance, this technique presents a design direction for camera-based perception pipelines, with expectations for further robustness through future multi-sensor fusion with Radar and LiDAR.