Scholarlink

<p>Accurately segmenting objects is essential for autonomous delivery vehicles to navigate safely and avoid obstacles in complex urban environments. In this work, we introduce a novel method for creating high-precision segmented Bird’s Eye View (BEV) maps from four surrounding images, aimed at supporting reliable operation to navigate safely. Our model includes a custom-designed LiDAR encoder that extracts multi-layer spatial features, enhancing depth-aware perception and improving the model’s ability to recognize small-scale objects like vehicles and other obstacles critical for navigation. Unlike conventional approaches that rely solely on image-based segmentation, our LiDAR encoder provides fine-grained spatial representations, ensuring robustness in challenging conditions. For classes like vehicles, which are relatively small in the context of the entire map, the IoU increased by approximately 7.92%, solely due to the integration of the LiDAR encoder. To support this work, we collected a diverse and high-resolution dataset with both LiDAR and image data, ensuring robust generalization across varied urban scenarios. Additionally, an attention mechanism is integrated to help the model focus on important regions across images, enhancing the segmentation accuracy across the entire scene. By combining LiDAR’s depth precision with an attention-driven refinement strategy, our hybrid LiDAR-attention framework enables more accurate and context-aware segmentation of objects in BEV representation. Compared to the baseline method, our hybrid LiDAR-attention framework enables more precise segmentation of objects in BEV representation. Our results show that this approach significantly improves small object recognition and boosts overall performance, offering a reliable solution for BEV segmentation in autonomous delivery applications.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Enhancing bird’s eye view segmentation with LiDAR encoders and attentive feature fusion

  • 期刊论文

  • 发表日期：

    2025年9月3日

查看全文
  • Sabir Hossain,
  • Xianke Lin

摘要

Accurately segmenting objects is essential for autonomous delivery vehicles to navigate safely and avoid obstacles in complex urban environments. In this work, we introduce a novel method for creating high-precision segmented Bird’s Eye View (BEV) maps from four surrounding images, aimed at supporting reliable operation to navigate safely. Our model includes a custom-designed LiDAR encoder that extracts multi-layer spatial features, enhancing depth-aware perception and improving the model’s ability to recognize small-scale objects like vehicles and other obstacles critical for navigation. Unlike conventional approaches that rely solely on image-based segmentation, our LiDAR encoder provides fine-grained spatial representations, ensuring robustness in challenging conditions. For classes like vehicles, which are relatively small in the context of the entire map, the IoU increased by approximately 7.92%, solely due to the integration of the LiDAR encoder. To support this work, we collected a diverse and high-resolution dataset with both LiDAR and image data, ensuring robust generalization across varied urban scenarios. Additionally, an attention mechanism is integrated to help the model focus on important regions across images, enhancing the segmentation accuracy across the entire scene. By combining LiDAR’s depth precision with an attention-driven refinement strategy, our hybrid LiDAR-attention framework enables more accurate and context-aware segmentation of objects in BEV representation. Compared to the baseline method, our hybrid LiDAR-attention framework enables more precise segmentation of objects in BEV representation. Our results show that this approach significantly improves small object recognition and boosts overall performance, offering a reliable solution for BEV segmentation in autonomous delivery applications.