基于场景自适应特征的行人重识别技术  

Pedestrian Reidentification Technology Based on Scenario Adaptive Features

作　　者：丁正彦 尚岩峰[1] DING Zhengyan;SHANG Yanfeng(Research Center on Internet of Things,the Third Research Institute of the Ministry of Public Security,Shanghai 201204,China)

机构地区：[1]公安部第三研究所,物联网技术研发中心,上海201204

出　　处：《微型电脑应用》2025年第2期17-19,共3页Microcomputer Applications

基　　金：上海市青年科技英才扬帆计划(20YF1409300)。

摘　　要：行人重识别是视频监控系统的关键技术之一,但现有基于深度网络模型的特征提取方法难以兼顾多种业务场景。针对上述问题,基于人脸识别数据集训练大尺度时空场景下的特征提取模型,同时基于人体识别数据集训练小尺度时空场景下的特征提取模型。采用知识蒸馏框架进行跨尺度多特征融合,并引入目标质量评价和人机协同约束机制实现多场景特征的自适应选择。实验结果表明,提出的基于场景自适应特征的行人重识别技术能够显著提升系统性能,通过实现多特征自适应融合,行人目标检索的平均准确率与单独使用人脸特征相比提升了55.94个百分点,与单独使用人体特征相比提升了14.56个百分点。此外,针对特征关注区域进行了可视化分析,进一步验证了不同尺度时空场景下行人特征之间存在的互补关系。Pedestrian reidentification is one of the key technologies in video surveillance system.However,the existing feature extraction method based on deep network model is difficult to cover different service scenarios.Aimed at the above problem,this paper trains the model in large-scale spatio-temporal scenarios based on the face indentification dataset,and trains the model in small-scale spatio-temporal scenarios based on the body indentification dataset at the same time.The knowledge distillation framework is used for cross-scale multiple features fusion,and the target quality evaluation and man-machine collaborative constraint mechanism are introduced to realize the adaptive selection of multiple scenario features.The experimental results indicate that the pedestrian reidentification technology based on scenario adaptive features can significantly improve the system performance.By realizing the adaptive fusion of multiple features,the average accuracy of pedestrian target retrieval increases by 55.94 percentage points compared with the face features,and increases by 14.56 percentage points compared with the body features.In addition,the visual analysis of feature-concerned areas is carried out to further verify the complementary relationship between pedestrian features in different scales and spatio-temporal ranges.

关 键 词：行人重识别 场景自适应 跨尺度多特征融合 互补关系 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]