摘要
文中提出一种快速构建异常行为检测数据集方法,该方法以一种半自动的方式完成数据集的构建,有助于减少构建过程中人工操作的工作量。首先以网络爬虫的方式自动地从互联网上搜索并下载包含指定动作的视频,之后以当前SOTA(state-of-the-art)的目标检测模型作为人物空间位置检测器,最后以人工标注和行为检测模型相结合的迭代方式完成人物行为的标注,其中需要手工完成的主要包括对下载的视频的挑选、人物边框核对以及一部分的行为标注,手工部分的工作量仅占整个任务的工作量的15%左右。实验表明,由该方法所构建的数据集可以作为异常行为检测模型的训练集使用,验证了该方法的有效性。通过该方法可以快速地构建一个大尺度、高质量的行为检测数据集,将有助于推动异常行为检测研究的深入开展。
We propose a fast strategy to build abnormal action detection dataset.This strategy completes the construction of dataset in a semi-automatic way,which significantly reduces the workloads of manual operation in the construction pipeline.This method searches and downloads videos containing specified actions automatically from the Internet in the form of web crawler.The object detection model with the state-of-the-art performance is used as the person spatial detector.Finally,the human action annotations are completed iteratively by combining manual annotation and action detection model.What needs to be done manually is the selection of downloaded video,the check of person bounding boxes and a part of action annotations.In the experiment section,it demonstrates that the dataset constructed by this method is effective and can be used as the training set of abnormal action detection model.Furthermore,it shows that this method can quickly build a large-scale and high-quality action detection dataset,which will accelerate the development of research and application of abnormal action detection.
作者
杜潘飞
王志辉
李雄伟
朱永旺
DU Pan-fei;WANG Zhi-hui;LI Xiong-wei;ZHU Yong-wang(Shijiazhuang Campus,The Army Engineering University of PLA,Shijiazhuang 050003,China;Hebei Construction&Investment Group Co.,Ltd.,Shijiazhuang 050001,China)
出处
《计算机技术与发展》
2021年第9期155-160,共6页
Computer Technology and Development
基金
国家青年科学基金(61602505)。
关键词
数据集构建
行为识别
目标检测
半自动构建方法
异常行为
dataset construction
action recognition
object detection
semi-automatic construction method
abnormal action
作者简介
通讯作者:杜潘飞(1991-),男,硕士研究生,研究方向为计算机视觉;李雄伟,教授,硕导,博士,研究方向为信息安全。
