HP-BEV: a bird’s-eye view segmentation model based on spatial height modeling

For autonomous driving tasks, obtaining information about targets in the vehicle coordinate system is essential for downstream tasks such as localization, prediction, and planning. Currently, forward projection-based bird’s-eye view (BEV) algorithms are heavily reliant on depth estimation, which consumes significant computational resources. By contrast, backward projection can result in inaccurate correspondences between front-view and BEV, and research on height modeling in BEV space remains limited. To address this challenge, we propose an end-to-end BEV segmentation network, HPBEV, which explicitly models the height in BEV space to ensure the accuracy of backward projection. Furthermore, to maintain the model’s real-time performance, we introduce a cross-scale cross-attention module. By utilizing smaller-sized BEV queries as input and upsampling the attention weights, this module effectively replaces the use of larger-sized BEV Queries, enabling the BEV space to incorporate richer and more detailed front-view feature information. The proposed model operates in real-time on an RTX 3090 GPU and demonstrates certain performance improvements over existing models.