Work

ウェアラブルセンサを用いた人間行動認識のための高効率な一次元信号処理手法

Year: 2026
Type: dissertation
Abstract: Human activity recognition (HAR) aims to perform real-time monitoring of target users' activity status using data from various sensors.According to the input data type, we can divide HAR tasks into the following categories: visionbased HAR, ambient-sensor-based HAR, and wearable-sensor-based HAR, among which the wearable-sensor-based HAR has the advantages of requiring fewer computational resources and environments, being more robust to noise, and being more privacy-friendly, making it promising and popular in the research areas, as well as the applications, including activity monitoring, surveillance systems, healthcare, and human-computer interaction.As a typical classification problem, a wearable-sensor-based HAR task usually consists of the following steps: data collection, data preprocessing, data representation, and modeling.Traditionally the input data, in most cases accelerometer signals and gyroscope collected from wearable sensors, will first be preprocessed to reduce the influence of irrelevant factors in the data, then we apply a sliding window to obtain the signal frames at each time step, and the features can be extracted based on these signal frames.Finally, we input the feature vector to the classification models, and the model will output the most likely activity.In HAR research, designing an efficient system has always been the most critical topic and a challenging problem.The enhancements of steps in the HAR pipeline are considered as the main methods in previous research, such as more efficient data representation and new designs of models with higher performance.In this dissertation, we propose 3 studies to research improving the accuracy of wearable-sensor-based HAR via data processing techniques from 3 different directions.Specifically, we propose: (1) SDT-DA, a novel Signal Decomposition and Transform based Data Augmentation method.DA is an effective preprocessing method to address the overfitting problem caused by data insufficiency, while data augmentation for sensor signals has v As I approach the conclusion of my three-year academic journey in Japan, I would like to express my sincere gratitude to the following individuals, whose support and guidance have enabled me to reach this point.First, I would like to express my sincere gratitude
Author: 李, 鋼凱
Institution: Kyushu University
Language: English
Open Access status: green
OpenAlex ID: W7225277992
Cites: 84
Cited by:
Related to: 10