DSWHAR: A Dynamic Sliding Window Based Human Activity Recognition Method

Sensory data has been widely used for human activity recognition (HAR), where sliding window (SW) is one of the typical methods to segment continuous signals. Most existing HAR methods select fixed-length sliding windows for different activities. However, with complex and diverse human activities, a single fixed sliding window is not suitable for acquiring features of different activities. Some studies have dynamically adjusted the window size based on the activity of feature difference segmentation, but the continuous occurrence of the activity can lead to an oversized window for segmentation and easily introduce additional noise leading to degradation of classification performance. To tackle the above problem, we propose a novel dynamic sliding window based human activity recognition method, called DSWHAR. In the training phase, DSWHAR firstly explores the relationship between the optimal sliding window length and the signal spectrum of each target activity, then a candidate window set is built for the target activity and multiple classifiers are trained using different windows. During the recognition phase, the latest prediction probability is used to dynamically select a suitable window for the next prediction, and the corresponding classifier is used for the final recognition. We compare DSWHAR with the method with fixed sliding windows and dynamic sliding window algorithm on the MHEALTH dataset, respectively, and the experimental results show that DSWHAR not only achieves better overall HAR accuracy but also has unbiased performance on different activities.