A detailed human activity transition recognition framework for grossly labeled data from smartphone accelerometer

Smartphone based human activity monitoring and recognition play an important role in several medical applications, such as eldercare, diabetic patient monitoring, post-trauma recovery after surgery. However, it is more important to recognize the activity sequences in terms of transitions. In this work, we have designed a detailed activity transition recognition framework that can identify a set of activity transitions and their sequence for a time window. This enables us to extract more meaningful insight about the subject’s physical and behavioral context. However, precise labeling of training data for detailed activity transitions at every time instance is required for this purpose. But, due to non uniformity of individual gait, the labeling tends to be error prone. Accordingly, our contribution in this work is to formulate the activity transition detection problem as a multiple instance learning problem to deal with imprecise labeling of data. The proposed human activity transition recognition framework forms an ensemble model based on different MIML-kNN distance metrics. The ensemble model helps to find both the activity sequence as well as multiple activity transition. The framework is implemented for a real dataset collected from 8 users. It is found to be working adequately (average precision 0.94).