Scholarlink

<p>Pedestrian re-identification (Re-ID) is an important task in intelligent surveillance and public safety. Traditional pedestrian re-identification methods show obvious limitations when facing the occlusion problem, leading to a significant decrease in re-identification accuracy. For this reason, we design an occlusion perceptual attention module (OPAM), which seeks to improve the model’s capacity to grasp both local and global contextual information. Secondly, we bring in an improved feature fusion module FtF (Feature to Feature), which aims to fully utilize the rich information of convolutional features to enhance the feature representation capability of the visual transformer model. Finally, this paper constructs a comprehensive loss function robust triplet loss (RTL), which combines the triad loss and occlusion perception loss to enhance the performance and efficiency of pedestrian re-recognition. We conduct experiments on two recognized occluded pedestrian datasets, with Rank-1 of 73.6% and mAP of 63.2% on the Occluded-Duke dataset, which is an increase of 2.6% and 2.2% from baseline, respectively; and Rank-1 of 90.8% and mAP of 82.4% on the DukeMTMC-reID dataset. The good performance compared with state-of-the-art methods fully validates its validity and generalizability.</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

Research on occlusion pedestrian re-identification based on ViT model

  • 期刊论文

  • 发表日期：

    2025年1月3日

查看全文
  • Yuepeng Guo,
  • ZhenPing Lan,
  • Yanguo Sun,
  • Yuheng Sun,
  • Xinxin Li,
  • Yuru Wang,
  • Bo Li

摘要

Pedestrian re-identification (Re-ID) is an important task in intelligent surveillance and public safety. Traditional pedestrian re-identification methods show obvious limitations when facing the occlusion problem, leading to a significant decrease in re-identification accuracy. For this reason, we design an occlusion perceptual attention module (OPAM), which seeks to improve the model’s capacity to grasp both local and global contextual information. Secondly, we bring in an improved feature fusion module FtF (Feature to Feature), which aims to fully utilize the rich information of convolutional features to enhance the feature representation capability of the visual transformer model. Finally, this paper constructs a comprehensive loss function robust triplet loss (RTL), which combines the triad loss and occlusion perception loss to enhance the performance and efficiency of pedestrian re-recognition. We conduct experiments on two recognized occluded pedestrian datasets, with Rank-1 of 73.6% and mAP of 63.2% on the Occluded-Duke dataset, which is an increase of 2.6% and 2.2% from baseline, respectively; and Rank-1 of 90.8% and mAP of 82.4% on the DukeMTMC-reID dataset. The good performance compared with state-of-the-art methods fully validates its validity and generalizability.