Monitoring and Coaching the Use of Home Medical Devices

Despite the popularity of home medical devices, serious safety concerns have been raised, because the use-errors of home medical devices have linked to a large number of fatal hazards. To resolve the problem, we introduce a cognitive assistive system to automatically monitor the use of home medical devices. Being able to accurately recognize user operations is one of the most important functionalities of the proposed system. However, even though various action recognition algorithms have been proposed in recent years, it is still unknown whether they are adequate for recognizing operations in using home medical devices. Since the lack of the corresponding database is the main reason causing the situation, at the first part of this paper, we present a database specially designed for studying the use of home medical devices. Then, we evaluate the performance of the existing approaches on the proposed database. Although using state-of-art approaches which have demonstrated near perfect performance in recognizing certain general human actions, we observe significant performance drop when applying it to recognize device operations. We conclude that the tiny actions involved in using devices is one of the most important reasons leading to the performance decrease. To accurately recognize tiny actions, it’s critical to focus on where the target action happens, namely the region of interest (ROI) and have more elaborate action modeling based on the ROI. Therefore, in the second part of this paper, we introduce a simple but effective approach to estimating ROI for recognizing tiny actions. The key idea of this method is to analyze the correlation between an action and the sub-regions of a frame. The estimated ROI is then used as a filter for building more accurate action representations. Experimental results show significant performance improvements over the baseline methods by using the estimated ROI for action recognition. We also introduce an interaction framework, which considers both the confidence of the detection as well as the seriousness of the potential error with messages to the user that take both aspects into account.