FLAAP: An open Human Activity Recognition (HAR) dataset for learning and finding the associated activity patterns

The purpose of this research is to propose a dataset for Human Activity Recognition (HAR) utilizing smartphone sensors. The dataset consists of 10 activities that were completed throughout three trials by eight different people. After examining the results of each experiment, the dataset only contains one of the optimal activity patterns. For intelligent learning algorithms to detect and understand the related activity patterns, this time-series data collected from smartphone-embedded (accelerometer and gyroscope) sensors provided valuable information. With the growing need for context-aware apps relating to people depending on their activities such as sports, health care, surveillance, yoga, the gym, and many more. The creation of solutions for these kinds of problems depends on the availability of such data. To collect data while doing the human activities indicated in this article, we give a HAR dataset of measurements taken from smartphone sensors targeted at the subject's body. Millions of raw sensor activity data samples were continuously collected between February 1st and May 31st of 2022 at sampling speeds of 100 Hz. The performance evaluation of various machine and deep learning algorithms for the learning and identification of activity patterns might benefit from this dataset. The performance of the learning algorithms while using various data preprocessing methods, knowledge transfer to target domains, and other methods may be of special interest to the research community. Such learning algorithms may also be used with HAR data that has been gathered over a long period. This makes it possible to determine activity patterns for a certain period. After that, after a predetermined period, we can detect changes. For patients with dementia, the physically challenged, the elderly, and youngsters who care for others, among others, this can assist in recognizing abnormalities and delivering early therapies.

本研究旨在构建一款适用于人类活动识别（Human Activity Recognition, HAR）的智能手机传感器数据集。该数据集包含8名受试者完成3轮试验所采集的10类活动数据，经校验各试验结果后，数据集仅留存各类活动的最优活动模式样本。针对用于检测与识别相关活动模式的智能学习算法而言，本数据集采用智能手机内置加速度计（accelerometer）与陀螺仪（gyroscope）采集的时序数据，可为其提供极具价值的参考依据。随着基于用户活动的情境感知应用需求日益增长，此类应用覆盖运动、医疗保健、安防监测、瑜伽、健身等诸多场景，针对上述场景的问题求解方案，其构建离不开此类数据集的支撑。为采集本文所述人类活动过程中的数据，我们构建了一款基于智能手机传感器、针对受试者身体活动的HAR数据集，该数据集于2022年2月1日至5月31日期间，以100Hz的采样频率持续采集了数百万条原始传感器活动数据样本。该数据集可用于评估各类机器学习与深度学习算法在活动模式学习与识别任务中的性能表现，研究人员可基于本数据集探究各类数据预处理方法、跨域知识迁移等技术对学习算法性能的影响，此方向或为学术界重点关注的研究议题。此类学习算法亦可适配长期采集的HAR时序数据，从而实现特定时段内活动模式的识别，并可在预设时段后检测活动模式的变化情况。本数据集可辅助识别痴呆患者、残障人士、老年人及照护者等群体的活动异常，并助力开展早期干预治疗。