基于迁移学习的行人检测研究进展  

Review of pedestrian detection based on transfer learning

作　　者：邵松[1,2] 刘宏[1,2] 王向东[1,2] 钱跃良[1,2] 

机构地区：[1]中国科学院计算技术研究所普适计算研究中心,北京100190 [2]中国科学院计算技术研究所移动计算与新型终端北京市重点实验室,北京100190

出　　处：《计算机工程与应用》2014年第24期156-163,共8页Computer Engineering and Applications

基　　金：国家自然科学基金(No.61202209);北京市自然科学基金(No.4142051)

摘　　要：行人检测是计算机视觉的研究热点和难点,近年来基于机器学习的行人检测技术取得了长足的进步,但由于不同场景的数据分布存在差异,已有检测器在新场景下的行人检测性能出现显著下降。为了避免繁琐的人工标注,充分利用原有检测器和标注样本,基于迁移学习的行人检测研究受到越来越多的关注。对其中涉及到的样本获取、迁移学习机制等关键技术进行综述,并从多个角度对现有方法进行分析和比较,最后对该技术的未来进行展望。Pedestrian detection is an active area of research with challenge in computer vision. In recent years, pedestrian detection based on machine learning has achieved great development. However, since data of various application scenes are under different data distributions, the performance of a well-trained detector drops significantly in a new scene. In order to avoid the effort of manual labeling and make full use of the original detector and labeled samples, pedestrian detection based on transfer learning has attracted more and more attention. This paper reviews pedestrian detection based on transfer learning, involving sample collection, transfer learning and detector optimization. Recent research on this topic is summarized and compared in different ways. Future directions are discussed.

关 键 词：行人检测 检测器 迁移学习 场景自适应 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]