A Decision Level Fusion and Signal Analysis Technique for Activity Segmentation and Recognition on Smart Phones

The objective of this work is to recognize modes of locomotion and transportation accurately, with special emphasis on precise detection of transitions between different activities. The recognition of activities of daily living (ADLs), specifically modes of locomotion and transportation, provides an important context for many ubiquitous sensing applications. The precise detection of activity transition time is also important for applications that require immediate response. Many prior signal processing techniques use a fixed-length window for signal segmentation, which leads to poor performance for detecting activity transitions due to the limitation of a single window size. In this paper, we construct weak classifiers based on different window sizes and propose a decision level fusion approach to effectively classify and assign a label for each sample by fusing the decisions from all weak classifiers. Moreover, we propose a set of phone orientation independent features to ensure the system can work with arbitrary phone orientation. Our team, The Drifters, attained an F-score improvement of 1.9%, increasing from 94% to 95.9%, using our proposed method compared to using a single fixed-size window segmentation technique.