Scholarlink

<p>Reliable ground-plane localization of pedestrians within the 0–15&#xa0;m near-field zone is a prerequisite for safety-critical automotive functions, yet the fisheye cameras that dominate production surround-view systems introduce severe optical distortions that render conventional rectilinear models fundamentally inadequate. Monocular methods are further hampered by scale ambiguity arising from variable pedestrian stature, while end-to-end architectures exhibit feature degradation at the image periphery due to non-uniform radial distortion. A modular multi-camera framework is introduced to address these limitations through three key components: (i) a fisheye-aware Virtual Camera (VCAM) formulation that mitigates distortion via object-centric rotation, (ii) a weighted foot-projection strategy that uses multi-camera geometry for height-agnostic ground-plane localization in bird’s-eye view, and (iii) a vector-based geometric intersection engine that replaces rasterized occupancy checks, yielding a <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(3000\times\)</EquationSource> </InlineEquation> reduction in perception latency. Evaluated on the Valeo Near-Field (VNF) dataset, the framework achieves a Mean Absolute Error (MAE) of 0.55&#xa0;m in the 0–5&#xa0;m range—a 2.5<InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(\times\)</EquationSource> </InlineEquation> improvement over height-based baselines. End-to-end BEV architectures fail to converge on fisheye data, yielding MAE &gt;2.5&#xa0;m versus the proposed method, confirming the framework’s superior robustness under fisheye distortion. These results establish the framework’s suitability for production-oriented surround-view systems and its alignment with near-field safety requirements under ISO&#xa0;21448 (SOTIF).</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

Fisheye-aware multi-camera framework for automotive near-field pedestrian localization

  • 期刊论文

  • 发表日期：

    2026年7月17日

查看全文
  • Basem Barakat,
  • Muhammad Islam,
  • Mohammed Rasmy,
  • Elsayed Hemayed,
  • Ahmed Radwan,
  • Antonyo Musabini

摘要

Reliable ground-plane localization of pedestrians within the 0–15 m near-field zone is a prerequisite for safety-critical automotive functions, yet the fisheye cameras that dominate production surround-view systems introduce severe optical distortions that render conventional rectilinear models fundamentally inadequate. Monocular methods are further hampered by scale ambiguity arising from variable pedestrian stature, while end-to-end architectures exhibit feature degradation at the image periphery due to non-uniform radial distortion. A modular multi-camera framework is introduced to address these limitations through three key components: (i) a fisheye-aware Virtual Camera (VCAM) formulation that mitigates distortion via object-centric rotation, (ii) a weighted foot-projection strategy that uses multi-camera geometry for height-agnostic ground-plane localization in bird’s-eye view, and (iii) a vector-based geometric intersection engine that replaces rasterized occupancy checks, yielding a \(3000\times\) reduction in perception latency. Evaluated on the Valeo Near-Field (VNF) dataset, the framework achieves a Mean Absolute Error (MAE) of 0.55 m in the 0–5 m range—a 2.5 \(\times\) improvement over height-based baselines. End-to-end BEV architectures fail to converge on fisheye data, yielding MAE >2.5 m versus the proposed method, confirming the framework’s superior robustness under fisheye distortion. These results establish the framework’s suitability for production-oriented surround-view systems and its alignment with near-field safety requirements under ISO 21448 (SOTIF).