Scholarlink

<p>Overhead fisheye cameras are widely used for pedestrian monitoring because of their wide field of view. However, dense crowds and severe radial distortion create pronounced scale imbalance and highly non-uniform object distributions, making real-time deployment particularly challenging. To address this issue, we revisit detector design from the perspective of scale reallocation. Instead of distributing computation uniformly across the image, the proposed method allocates more representational and optimization capacity to distortion-sensitive small pedestrians, while reducing redundant processing in well-covered regions. Specifically, we design an Efficient Scale Reallocation AFPN (ESR-AFPN) to strengthen high-resolution features and improve small-object perception with limited additional cost. We further introduce a wavelet downsampling convolution module (WDSConv) to suppress aliasing and preserve discriminative microstructures during resolution reduction. To better handle geometric irregularities, we develop FNA-IoU, a regression loss that emphasizes hard samples caused by distortion and rotation. In addition, a fisheye-oriented training strategy is adopted to improve robustness to orientation uncertainty. Extensive experiments on the LOAF dataset show that ESR-AFPN reduces the number of parameters by 9% and GFLOPs by 24% compared with AFPN while improving mAP@75 by about 1%. The complete detector achieves an 8.8% gain in mAP@75 over the baseline, including an approximately 13.9% improvement for small pedestrians. The proposed method maintains over 90 FPS, yielding a favorable accuracy–efficiency trade-off. Cross-dataset evaluation on FRIDA further reaches 92.8% mAP@50, demonstrating stable generalization capability. The code is available at: <a href="https://github.com/Dear2You/ESR.git">https://github.com/Dear2You/ESR.git</a>.</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

Efficient scale reallocation for rotated pedestrian detection in overhead fisheye images

  • 期刊论文

  • 发表日期：

    2026年5月17日

查看全文
  • Chenyu Wang,
  • Wenhai Dong,
  • Renjie Qiao,
  • Chengtao Cai

摘要

Overhead fisheye cameras are widely used for pedestrian monitoring because of their wide field of view. However, dense crowds and severe radial distortion create pronounced scale imbalance and highly non-uniform object distributions, making real-time deployment particularly challenging. To address this issue, we revisit detector design from the perspective of scale reallocation. Instead of distributing computation uniformly across the image, the proposed method allocates more representational and optimization capacity to distortion-sensitive small pedestrians, while reducing redundant processing in well-covered regions. Specifically, we design an Efficient Scale Reallocation AFPN (ESR-AFPN) to strengthen high-resolution features and improve small-object perception with limited additional cost. We further introduce a wavelet downsampling convolution module (WDSConv) to suppress aliasing and preserve discriminative microstructures during resolution reduction. To better handle geometric irregularities, we develop FNA-IoU, a regression loss that emphasizes hard samples caused by distortion and rotation. In addition, a fisheye-oriented training strategy is adopted to improve robustness to orientation uncertainty. Extensive experiments on the LOAF dataset show that ESR-AFPN reduces the number of parameters by 9% and GFLOPs by 24% compared with AFPN while improving mAP@75 by about 1%. The complete detector achieves an 8.8% gain in mAP@75 over the baseline, including an approximately 13.9% improvement for small pedestrians. The proposed method maintains over 90 FPS, yielding a favorable accuracy–efficiency trade-off. Cross-dataset evaluation on FRIDA further reaches 92.8% mAP@50, demonstrating stable generalization capability. The code is available at: https://github.com/Dear2You/ESR.git.