Hierarchical Signal Segmentation and Classification for Accurate Activity Recognition

The objective of this work is to determine various modes of locomotion and in particular identify the transition time from one mode of locomotion to another as accurately as possible. Recognizing human daily activities, specifically modes of locomotion and transportation, with smartphones provides important contextual insight that can enhance the effectiveness of many mobile applications. In particular, determining any transition from one mode of operation to another empowers applications to react in a timely manner to this contextual insight. Previous studies on activity recognition have utilized various fixed window sizes for signal segmentation and feature extraction. While extracting features from larger window size provides richer information to classifiers, it increases misclassification rate when a transition occurs in the middle of windows as the classifier assigns only one label to all samples within a window. This paper proposes a hierarchical signal segmentation approach to deal with the problem of fixed-size windows. This process begins by extracting a rich set of features from large segments of signal and predicting the activity. Segments that are suspected to contain more than one activity are then detected and split into smaller subwindows in order to fine-tune the label assignment. The search space of the classifier is narrowed down based on the initial estimation of the activity, and labels are assigned to each sub-window. Experimental results show that the proposed method improves the F1-score by 2% compared to using fixed windows for data segmentation. The paper presents the techniques employed in our team's (The Drifters) submission to the SHL recognition challenge.