Detecting Harmful Hand Behavior with Machine Learning from Wearable Motion Sensor Data
In medical care and special needs areas, human activity recognition helps doctors track the patients while they are unsupervised. In this paper, we will present our classifier system for detecting harmful hand behavior. The data comes from a wearable sensor on the user’s wrist. It collects signals in the three axes x, y and z. For each axis, it contains multiple attributes. Because reducing irrelevant attributes can decrease the time complexity and increase the accuracy, we started processing the raw data by ignoring some attributes from the whole attributes set. Our design approach is to apply a classification algorithm which generate an initial output, and then use a sequence post process to correct potentially incorrect initial outputs. The basic classifier algorithm is a decision tree, where we adopt the random forest approach to reduce generation error. Experimental results show that the system can get a 96% accuracy rate in detecting harmful behavior, and it can also obtain 95% accuracy rate distinguishing the ambiguous behaviors from the harmful behaviors.