Scholarlink

With the advancement of autonomous driving technology and FPGA-based gas pedals, deploying deep learning models on FPGAs has become crucial for rapidly delineating the alert area during train operations. First, we deployed FPN + ResNet18 semantic segmentation model in ZCU104 board to solve the problems of low energy efficiency ratio and poor device stability of deploying deep learning models in CPU and GPU. The model achieved an accuracy of 96.1% in detecting alert region on the test set, with a loss value of 0.1 and an mIoU of 72.4%, while maintaining a frame rate of 14.7 fps. Compared to models on CPU and GPU platforms, its energy efficiency ratio was improved by 143 times and 3.38 times, respectively. Then, we proposed a main orbit region detection method based on improved Hough transform to address the issues of low frame rate of deep learning model detection and high resource occupation rate of traditional machine vision techniques, which uses the semantic segmentation results as the key frame data for alert region detection, and performs well in linear track regions. The detection frame rate and the PA value of this method is 32.3 fps and 97.83%, which is 2.19 and 1.05 times higher than the frame rate and PA value of the previous semantic segmentation model, respectively. Moreover, the resource occupation of this model in the ZCU104 development board is almost the same as the previous model. Finally, we tested the overall project and obtained better results for the guarded area of the train track.

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

Design of Railway Track Warning Area Detection Accelerator Based on FPGA

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Jun Fu,
  • Baohua Wang,
  • Yan Zhou,
  • Kaiyu Zhang,
  • Yaqi Mi

摘要

With the advancement of autonomous driving technology and FPGA-based gas pedals, deploying deep learning models on FPGAs has become crucial for rapidly delineating the alert area during train operations. First, we deployed FPN + ResNet18 semantic segmentation model in ZCU104 board to solve the problems of low energy efficiency ratio and poor device stability of deploying deep learning models in CPU and GPU. The model achieved an accuracy of 96.1% in detecting alert region on the test set, with a loss value of 0.1 and an mIoU of 72.4%, while maintaining a frame rate of 14.7 fps. Compared to models on CPU and GPU platforms, its energy efficiency ratio was improved by 143 times and 3.38 times, respectively. Then, we proposed a main orbit region detection method based on improved Hough transform to address the issues of low frame rate of deep learning model detection and high resource occupation rate of traditional machine vision techniques, which uses the semantic segmentation results as the key frame data for alert region detection, and performs well in linear track regions. The detection frame rate and the PA value of this method is 32.3 fps and 97.83%, which is 2.19 and 1.05 times higher than the frame rate and PA value of the previous semantic segmentation model, respectively. Moreover, the resource occupation of this model in the ZCU104 development board is almost the same as the previous model. Finally, we tested the overall project and obtained better results for the guarded area of the train track.