Abstract: 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.

Key words: pedestrian detection, detector, transfer learning, domain adaptation

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。

关键词: 行人检测, 检测器, 迁移学习, 场景自适应