Improving Bag-Of-Words: Capturing Local Information for Motion-Based Activity Recognition.

Bag-of-Words (BoW) is one of the important techniques for activity recognition. Instead of dividing a continuous sensor streams into sliding windows with fixed time duration, it builds activity recognition models using histograms of primitive motion symbols. However, this BoW method losses the sequential information in the symbol sequences and limits the performance of activity recognition. In this paper, we propose an activity recognition approach to get rid of this limitation and consider longer time dependency by capturing local features from the symbol sequences. We use a set of small sliding windows inside the symbol sequences to capture local features. Our algorithm utilizes the physical knowledge where the sequence of the selected window size of symbols reflects the context and order of an activity. We evaluate the activity recognition approaches on two public datasets. The results show that our approach achieved stable improvement on all the datasets, compared with traditional statistical and BoW approaches.