Human Activity Recognition Through Deep Learning: Leveraging Unique and Common Feature Fusion in Wearable Multi-Sensor Systems

Abstract

Multi-sensor fusion plays a crucial role in human activity recognition (HAR) as it integrates diverse information from different sensors. However, determining the most significant features for classification and assigning appropriate weights to individual sensors can be a complex task. Additionally, the disparate data structures of these sensors pose challenges in finding a unified format for the effective fusion of heterogeneous data. To that end, this paper introduces a method that focuses on the fusion of unique and common features in wearable multi-sensor systems for human activity recognition, namely, UC Fusion. It presents a fusion method by merging the unique feature of each sensor and the common features of all the sensors. Besides, it also addresses the challenge of handling heterogeneous data by unifying the data format through segmentation and dimensional transformation. Extensive experiments were conducted on the UCI HAR dataset to evaluate the performance of UC Fusion. The results demonstrated that our proposed method achieved an average recognition accuracy of 96.84%. Furthermore, ablation studies were performed on each module of UC Fusion to assess their impact on the accuracy and the results confirmed the effectiveness of the proposed method.

Keywords: HAR, muti-sensor fusion, Deep Learning, common feature, unique feature