SOS: Real-time and accurate physical assault detection using smartphone

详细信息    查看全文

• 作者：Zehao Sun ; Shaojie Tang ; He Huang ; Zhenyu Zhu&#8230
• 关键词：Assault detection ; Smartphone sensing ; Activity recognition ; Machine learning
• 刊名：Peer-to-Peer Networking and Applications
• 出版年：2017
• 出版时间：March 2017
• 年：2017
• 卷：10
• 期：2
• 页码：395-410
• 全文大小：
• 刊物类别：Engineering
• 刊物主题：Communications Engineering, Networks; Information Systems and Communication Service; Computer Communication Networks; Signal,Image and Speech Processing;
• 出版者：Springer US
• ISSN：1936-6450
• 卷排序：10

文摘

Motivated by the reports about assaults on women, we take the first step to explore possibility of using off-the-shelf smartphone for physical assault detection. There are several kinds of crime offenses against persons, such as gunshot, battery, abuse, kidnapping and so on, which are distinguished by form, severity, duration, etc. In this paper, we aim at detecting those severe and non-instantaneous physical assaults using accelerometer in smartphone. We collected 100 surveillance videos involving aggravated assaults, and extract the the pattern of actions for an individual under assaulting. The most difficult one among challenges in our design is the extraordinary complexity and diversity of actions under assaulting, which lead to an extremely hard, if not impossible, to perform fine-grained recognition. To this end, we decide to focus on the intensity and irregularity characteristics of aggravated assaults, based on which several features from time domain and frequency domain are extracted. Moreover, we proposed a combinatorial classification scheme considering individuality of user’s ADLs (Activities of Daily Living) and universality of differences between ADLs and assaults to most people. The data we used for training and testing are collected from simulated aggravated assaults which are performed by our volunteers in controlled settings. Our experiment results showed that aggravated assaults could be distinguished with the majority of ADLs in our proposed feature space, and our proposed system could correctly detect most instances of aggravated assault (FNR = 11.75 %) with low false alarm rate (0.067 times per day) and short delay (6.89 s).