Scholarlink

<p>Person search aims to locate all pedestrians from a complete image and identify the target pedestrian, which plays an important role in the field of intelligent video surveillance. However, influence come from factors such as occlusion may cause the miss detection of the pedestrians in the original image and affecting cross-camera person search tasks greatly. Therefore, based on the DINO detection model, a method is proposed for jointly optimizing pedestrian detection and re-identification (DINOPS): a part attention mechanism is introduced to integrate global and local pedestrian features, reduce missed detection and improve the re-identification ability. Specifically, for detection task, parameter updating and loss calculation strategies are introduced in the model to strengthen the detection performance. For the re-identification task, the model adopts strategy of multi-scale feature fusion to concatenate pedestrian features that are extracted from multiple scale feature maps, thus obtaining richer pedestrian feature representations. Final experimental results show that DINOPS achieves a 6.6% increase in mAP and a 0.6% increase in top-1 on the PRW dataset compared with the current state-of-the-art (SOTA), demonstrating the effectiveness of the proposed approach in person search.</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

Joint detection and re-identification for occluded person search

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年7月2日

查看全文
  • Ye Li,
  • Shizhen Shuai,
  • Yunhan Zhou,
  • Binbin Deng,
  • Dongxing Zhang

摘要

Person search aims to locate all pedestrians from a complete image and identify the target pedestrian, which plays an important role in the field of intelligent video surveillance. However, influence come from factors such as occlusion may cause the miss detection of the pedestrians in the original image and affecting cross-camera person search tasks greatly. Therefore, based on the DINO detection model, a method is proposed for jointly optimizing pedestrian detection and re-identification (DINOPS): a part attention mechanism is introduced to integrate global and local pedestrian features, reduce missed detection and improve the re-identification ability. Specifically, for detection task, parameter updating and loss calculation strategies are introduced in the model to strengthen the detection performance. For the re-identification task, the model adopts strategy of multi-scale feature fusion to concatenate pedestrian features that are extracted from multiple scale feature maps, thus obtaining richer pedestrian feature representations. Final experimental results show that DINOPS achieves a 6.6% increase in mAP and a 0.6% increase in top-1 on the PRW dataset compared with the current state-of-the-art (SOTA), demonstrating the effectiveness of the proposed approach in person search.