SW-HMM: A Method for Evaluating Confidence of Smartphone-Based Activity Recognition

With the development of sensing technology, the smartphone-based human activity recognition has become a promising research field of great concern. Besides the recognition result itself, the confidence which indicates the probability one sample is correctly recognized is also a very important issue. Different from calculating confidence based on the spatial distribution of samples, we evaluate the confidence of recognized results in the view of time dimension and propose a novel method named Sliding-Window-based Hidden Markov Model (SW-HMM). In SW-HMM, the results are first divided into subsequences with fixed size, then the probability of each subsequence is calculated based on HMM, finally the confidence of each result is evaluated according to the probabilities of all the subsequences. To compare the performance with previous works, we introduce a new metric of result confidence, named confusion rate. We have made several experiments on real dataset, and the results show that the confusion rate of SW-HMM is 40% to 50% lower than other popular methods.