Scholarlink

Visually impaired people (VIP) have been using the tactile walking surface indicator (TWSI) as a guide tool, but it is unable to provide information about the surrounding environment. In addition, its occupancy and damage are visible everywhere, making it insufficient for VIP demands. In this paper, we first propose a Multi-Scale Group Fusion (MSGF) module, which efficiently fuses two feature maps with low computational overhead. Then we design the MSMD-YOLO Network, which incorporates both MSGF and Multidimensional Collaborative Attention (MCA) into the YOLO framework. To guide VIP, we constructed the dataset called Visual-assistance Traffic Accessibility Dataset(ViTAD). The MSMD-YOLO achieves 71.8% accuracy on ViTAD and 68.9% on PASCAL VOC 2012. This shows that our model not only guides the VIP, but also has the ability to generalize. Furthermore, the MSMD-YOLO has low computational complexity and high inference speed, achieving 108 FPS. Additionally, we accelerate the post-processing part of the model based on FPGA. We design the Parallel-NMS algorithm, this algorithm parallelizes the sorting and IoU computation of the NMS algorithm. On the KV260 FPGA, the Parallel-NMS algorithm executes at 0.049 ms, which is 14.3 times faster than the RTX 4060Ti.

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

MSMD-YOLO: An Object Detection Network for Visually Impaired People with FPGA Acceleration

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Zhuowen Duan,
  • Qionglu Ruan,
  • Junqi Zhang,
  • Wenjie Chen

摘要

Visually impaired people (VIP) have been using the tactile walking surface indicator (TWSI) as a guide tool, but it is unable to provide information about the surrounding environment. In addition, its occupancy and damage are visible everywhere, making it insufficient for VIP demands. In this paper, we first propose a Multi-Scale Group Fusion (MSGF) module, which efficiently fuses two feature maps with low computational overhead. Then we design the MSMD-YOLO Network, which incorporates both MSGF and Multidimensional Collaborative Attention (MCA) into the YOLO framework. To guide VIP, we constructed the dataset called Visual-assistance Traffic Accessibility Dataset(ViTAD). The MSMD-YOLO achieves 71.8% accuracy on ViTAD and 68.9% on PASCAL VOC 2012. This shows that our model not only guides the VIP, but also has the ability to generalize. Furthermore, the MSMD-YOLO has low computational complexity and high inference speed, achieving 108 FPS. Additionally, we accelerate the post-processing part of the model based on FPGA. We design the Parallel-NMS algorithm, this algorithm parallelizes the sorting and IoU computation of the NMS algorithm. On the KV260 FPGA, the Parallel-NMS algorithm executes at 0.049 ms, which is 14.3 times faster than the RTX 4060Ti.