An Exploratory Study of Sequence Alignment for Improved Sensor-Based Human Activity Recognition

Sequence alignment (SA) is a well-established technique in bioinformatics for analyzing deoxyribonucleic acid (DNA), ribonucleic acid (RNA), or protein sequences and identifying regions of similarity. The main goal of SA is to discover relationships between strings of data by deploying a series of heuristic or probabilistic methods to align a new string (e.g., DNA of a new species) with an existing string (DNA of a known species). SA has also been used sporadically in linguistics, social sciences, and finance. In this paper, the authors explore the prospect of coupling machine learning (ML) and SA to improve the output of human activity recognition (HAR) methods. In particular, several field experiments are conducted to collect heterogeneous human motion data via wearable sensors. Collected data is further mined using ML to identify sequences of activities performed in each experiment. Given the inaccuracy of sensor readings and the limitations of ML algorithms especially in handling datasets from complex human activities such as those performed by construction workers, it is expected that the resulting activity sequences not fully match actual activity sequences as observed in the field. To further clean up this inherent noise, SA is deployed to refine imperfections in the resulting activity sequences by manipulating the output of HAR and ultimately aligning noisy activity sequences with ground truth sequences. The outcome of this work is a systematic method to improve the reliability of HAR from sensor readings, which can benefit decision-making as related to task planning, resource management, productivity monitoring, and ergonomic assessment.