BehaviorSight: Privacy enhancing wearable system to detect health risk behaviors in real-time.

项目总结/摘要 危害健康的行为，如暴饮暴食、吸烟、饮酒和不坚持服药， 导致发病率和死亡率上升。为了跟踪和干预这些危害健康的行为， 临床医生传统上依赖于自我报告。然而，自我报告是不准确和有偏见的。因此我们不能 使用自我报告来验证自由生活条件下的健康风险行为。因此，一种自动化技术， 确认危害健康的行为是非常必要的。 随着可穿戴设备（例如，智能手表），自动监控物理 活动是可能的。然而，这些设备通常不提供任何视觉确认，这使得 核查在自由生活条件下开展的活动。摄像机可以捕获视点视频，因此可以 用作可穿戴设备，用于捕获视频，以视觉确认活动，包括健康风险行为。 这些记录可以帮助我们更好地了解健康风险行为。此外，视频信息可以 自动处理以确认和验证健康风险行为。 记录敏感内容和旁观者的视频涉及隐私和道德问题。目前 没有隐私保护摄像头可以自动检测健康风险行为，大多数人都是 不愿意戴摄像头而不会引起隐私问题。除了隐私问题，人们更喜欢 不显眼、小巧且不需要频繁充电的可穿戴设备。因此，一个隐私保护， 不显眼的可穿戴相机将增加可穿戴性。 红外（IR）传感器阵列具有提供独立温度读数的潜力，这允许 判断一个物体是近还是远。红外传感器阵列可以帮助只记录佩戴者和物体 近距离观察，同时过滤掉远处的物体。IR传感器阵列具有小的功率覆盖区，因此提供了 更长的电池寿命。我们的项目旨在开发一个隐私意识，不引人注目，可穿戴，行为检测 该平台将使人们能够真实的及时发现和干预危害健康的行为。 在这个项目中，我们将（1）开发可穿戴的行为检测设备，允许视觉确认， 加重了佩戴者的负担。该设备将通过红外传感器阵列数据增强RGB摄像头数据， 记录和自动行为检测。(2)我们将测试各种设计，以确定用户的可接受性 戴上这个装置然后，我们将测试各种图像处理技术和机器学习算法， 确定检测健康风险行为的最佳算法。(3)我们将把表现最好的 行为检测算法，以便它可以在开发的可穿戴设备上运行。通过行为检测 在可接受的可穿戴设备上运行的算法，真实的时间检测健康风险行为的能力将 成为现实。最终，我们的可穿戴设备将允许研究人员测试和应用适当的行为 无论何时发生危害健康的行为，都要进行真实的及时干预，而不是依赖自我报告。

Project Summary/Abstract Health-risk behaviors, such as overeating, smoking, consuming alcohol, and not adhering to medication, are responsible for increases in morbidity and mortality. To track and intervene during these health-risk behaviors, clinicians traditionally rely on self-reports. However, self-reports are inaccurate and biased. Therefore, we cannot use self-reports to validate health-risk behaviors in free-living conditions. Thus, an automated technique for validating health-risk behaviors is extremely necessary. With the growth and popularity of wearable devices (e.g., smartwatches), automatic monitoring of physical activity is possible. However, the devices often do not provide any visual confirmation, making it challenging to verify activities performed in free-living conditions. Cameras can capture point-of-view videos and can thus be used as a wearable device to capture videos for visual confirmation of activities, including health-risk behaviors. Such recordings can help us better understand health-risk behaviors. Additionally, video information can be automatically processed to confirm and validate health-risk behaviors. Recording videos of sensitive content and bystanders is associated with privacy and ethical concerns. Currently there is no privacy-preserving camera that can automatically detect health-risk behaviors, and most people are unwilling to wear cameras without raising privacy concerns. In addition to privacy concerns, people prefer wearables that are unobtrusive and small and that do not require frequent charging. Thus, a privacy-preserving, unobtrusive wearable camera would increase wearability. Infrared (IR) sensor arrays have the potential to provide independent temperature readings, which allows determining whether an object is near or far. The IR sensor array can help record only the wearer and objects near the wearer, while filtering out distant objects. IR sensor arrays have a small power footprint, thus providing longer battery life. Our project aims to develop a privacy-conscious, unobtrusive, wearable, behavior-detection platform that will make it possible to detect and intervene upon health-risk behaviors in real time. In this project, we will (1) develop the wearable behavior-detection device that allows visual confirmation without burdening the wearer. The device will augment RGB camera data with IR sensor array data for privacy-conscious recording and automatic behavior detection. (2) We will test various designs to determine a user's acceptability to wear the device. Then, we will test various image processing techniques and machine learning algorithms to determine the best algorithm for detecting health-risk behaviors. (3) We will incorporate the best-performing behavior-detection algorithm so that it can run on the developed wearable device. With a behavior-detection algorithm running on an acceptable wearable device, the ability to detect health-risk behaviors in real time will become a reality. Ultimately, our wearable device will allow researchers to test and apply appropriate behavioral interventions in real time, rather than relying on self-reports, whenever health-risk behaviors occur.