吕轩轩^1，2，宫法明^3*

1.内燃机可靠性国家重点实验室，山东 潍坊 261061；2.潍柴动力股份有限公司 软件研究院，山东 潍坊 261000；

3.中国石油大学（华东），山东 青岛  266000

摘要：针对传统的人工标注大量训练数据集存在的成本高、效率低、难以满足大型行人检测系统检测要求的问题，研究一种半自动标注行人的方法。以静止的单目摄像机拍摄的监控视频图像为对象，利用背景差分以及显著性检测方法将前景、背景分离，提取出前景目标。试验结果表明，相对于传统的人工标注数据集的方法该方法能提高数据集标注的效率，形成的数据库可以为行人检测系统提供数据。

关键词：行人检测；背景减法；显著性检测；深度学习；半自动数据集标注

Abstract: Pedestrian detection is a research hotspot in the field of computer vision. The role of national public security, traffic control and other fields cannot be ignored. The traditional method uses a convolutional neural network to train the detection model and manually label a large number of training datasets. This type of method greatly improves the detection accuracy, but repeated manual labeling is costly and inefficient, and it is difficult to meet the detection requirements for large pedestrian detection systems. Aiming at the surveillance video images taken by a stationary monocular camera, a method of semi-automatic annotation of pedestrians is proposed. This detection framework uses the video object segmentation method to separate the foreground background and extract the foreground object, so as to design a semi-automatic dataset annotation method. The experimental results show that the proposed method can improve the efficiency of data annotation compared with the traditional manual annotation dataset and can be applied to the real-time pedestrian detection system.

Keywords: pedestrian detection; background subtraction; saliency detection; deep learning; semi-automatic data annotation