CADDI: An in-Class Activity Detection Dataset using IMU data from low-cost sensors

Data DescriptionThe CADDI dataset is designed to support research in in-class activity recognition using IMU data from low-cost sensors. It provides multimodal data capturing 19 different activities performed by 12 participants in a classroom environment, utilizing both IMU sensors from a Samsung Galaxy Watch 5 and synchronized stereo camera images. This dataset enables the development and validation of activity recognition models using sensor fusion techniques.Data Generation ProceduresThe data collection process involved recording both continuous and instantaneous activities that typically occur in a classroom setting. The activities were captured using a custom setup, which included:A Samsung Galaxy Watch 5 to collect accelerometer, gyroscope, and rotation vector data at 100Hz.A ZED stereo camera capturing 1080p images at 25-30 fps.A synchronized computer acting as a data hub, receiving IMU data and storing images in real-time.A D-Link DSR-1000AC router for wireless communication between the smartwatch and the computer.Participants were instructed to arrange their workspace as they would in a real classroom, including a laptop, notebook, pens, and a backpack. Data collection was performed under realistic conditions, ensuring that activities were captured naturally.Temporal and Spatial ScopeThe dataset contains a total of 472.03 minutes of recorded data.The IMU sensors operate at 100Hz, while the stereo camera captures images at 25-30Hz.Data was collected from 12 participants, each performing all 19 activities multiple times.The geographical scope of data collection was Alicante, Spain, under controlled indoor conditions.Dataset ComponentsThe dataset is organized into JSON and PNG files, structured hierarchically:IMU Data: Stored in JSON files, containing:Samsung Linear Acceleration Sensor (X, Y, Z values, 100Hz)LSM6DSO Gyroscope (X, Y, Z values, 100Hz)Samsung Rotation Vector (X, Y, Z, W quaternion values, 100Hz)Samsung HR Sensor (heart rate, 1Hz)OPT3007 Light Sensor (ambient light levels, 5Hz)Stereo Camera Images: High-resolution 1920×1080 PNG files from left and right cameras.Synchronization: Each IMU data record and image is timestamped for precise alignment.Data StructureThe dataset is divided into continuous and instantaneous activities:Continuous Activities (e.g., typing, writing, drawing) were recorded for 210 seconds, with the central 200 seconds retained.Instantaneous Activities (e.g., raising a hand, drinking) were repeated 20 times per participant, with data captured only during execution.The dataset is structured as:/continuous/subject_id/activity_name/
/camera_a/ → Left camera images
/camera_b/ → Right camera images
/sensors/ → JSON files with IMU data

/instantaneous/subject_id/activity_name/repetition_id/
/camera_a/
/camera_b/
/sensors/
Data Quality & Missing DataThe smartwatch buffers 100 readings per second before sending them, ensuring minimal data loss.Synchronization latency between the smartwatch and the computer is negligible.Not all IMU samples have corresponding images due to different recording rates.Outliers and anomalies were handled by discarding incomplete sequences at the start and end of continuous activities.Error Ranges & LimitationsSensor data may contain noise due to minor hand movements.The heart rate sensor operates at 1Hz, limiting its temporal resolution.Camera exposure settings were automatically adjusted, which may introduce slight variations in lighting.File Formats & Software CompatibilityIMU data is stored in JSON format, readable with Python’s json library.Images are in PNG format, compatible with all standard image processing tools.Recommended libraries for data analysis:Python: numpy, pandas, scikit-learn, tensorflow, pytorchVisualization: matplotlib, seabornDeep Learning: Keras, PyTorchPotential ApplicationsDevelopment of activity recognition models in educational settings.Study of student engagement based on movement patterns.Investigation of sensor fusion techniques combining visual and IMU data.This dataset represents a unique contribution to activity recognition research, providing rich multimodal data for developing robust models in real-world educational environments.CitationIf you find this project helpful for your research, please cite our work using the following bibtex entry:@misc{marquezcarpintero2025caddiinclassactivitydetection,
   title={CADDI: An in-Class Activity Detection Dataset using IMU data from low-cost sensors}, 
   author={Luis Marquez-Carpintero and Sergio Suescun-Ferrandiz and Monica Pina-Navarro and Miguel Cazorla and Francisco Gomez-Donoso},
   year={2025},
   eprint={2503.02853},
   archivePrefix={arXiv},
   primaryClass={cs.CV},
   url={https://arxiv.org/abs/2503.02853}, 
}

数据集描述
CADDI数据集旨在支持基于低成本惯性测量单元（IMU）的课堂活动识别研究。该数据集提供多模态数据，涵盖12名参与者在课堂环境中完成的19种不同活动，同时采集了三星Galaxy Watch 5的IMU传感器数据与同步立体相机图像。本数据集可用于开发和验证基于传感器融合技术的活动识别模型。

## 数据采集流程
本次数据采集记录了课堂场景中常见的持续性与瞬时性活动，采用定制化设备搭建方案，包含以下设备：
1. 三星Galaxy Watch 5：以100Hz的频率采集加速度计、陀螺仪与旋转矢量数据；
2. ZED立体相机：以25-30 fps的帧率采集1080p分辨率图像；
3. 同步计算机：作为数据枢纽，实时接收IMU数据并存储图像；
4. D-Link DSR-1000AC路由器：用于智能手表与计算机间的无线通信。

参与者被要求按照真实课堂场景布置工作台，配备笔记本电脑、笔记本、笔与背包。数据采集在真实场景下进行，确保活动记录自然真实。

## 时空范围
本数据集总记录时长为472.03分钟。IMU传感器采样率为100Hz，立体相机图像采集帧率为25-30Hz。数据来自12名参与者，每名参与者重复完成全部19种活动多次。数据采集地点为西班牙阿利坎特，在受控室内环境下完成。

## 数据集组成
数据集以层级结构组织为JSON与PNG文件，具体组成如下：
### IMU数据
存储于JSON文件中，包含以下内容：
- 三星线性加速度传感器（X、Y、Z轴数值，100Hz）
- LSM6DSO陀螺仪（X、Y、Z轴数值，100Hz）
- 三星旋转矢量传感器（X、Y、Z、W四元数数值，100Hz）
- 三星心率传感器（心率数据，1Hz）
- OPT3007光线传感器（环境光照强度，5Hz）

### 立体相机图像
左右相机采集的1920×1080高分辨率PNG图像文件。
### 同步机制
每条IMU数据记录与图像均带有时间戳，可实现精准对齐。

## 数据结构
数据集分为持续性活动与瞬时性活动两类：
- 持续性活动（如打字、书写、绘图）的录制时长为210秒，仅保留中间200秒的数据。
- 瞬时性活动（如举手、饮水）每名参与者需重复完成20次，仅采集活动执行阶段的数据。

数据集的层级目录结构如下：
/连续活动/参与者ID/活动名称/
/camera_a/ → 左相机图像
/camera_b/ → 右相机图像
/sensors/ → 包含IMU数据的JSON文件

/瞬时活动/参与者ID/活动名称/重复次数ID/
/camera_a/
/camera_b/
/sensors/

## 数据质量与缺失数据处理
智能手表每秒钟缓存100条读数后再发送，最大限度减少数据丢失。智能手表与计算机间的同步延迟可忽略不计。由于采集帧率不同，并非所有IMU样本都有对应的图像。针对异常值与异常数据，通过丢弃持续性活动首尾的不完整序列进行处理。

## 误差范围与局限性
传感器数据可能因手部轻微移动产生噪声。心率传感器采样率为1Hz，限制了其时间分辨率。相机曝光设置为自动调整，可能会引入轻微的光照变化。

## 文件格式与软件兼容性
IMU数据以JSON格式存储，可通过Python的json库读取。图像采用PNG格式，兼容所有标准图像处理工具。
推荐用于数据分析的工具库：
- Python生态：numpy、pandas、scikit-learn、tensorflow、pytorch
- 可视化工具：matplotlib、seaborn
- 深度学习框架：Keras、PyTorch

## 潜在应用场景
1. 教育场景下的活动识别模型开发
2. 基于运动模式的学生参与度研究
3. 结合视觉与IMU数据的传感器融合技术研究

本数据集为活动识别研究提供了独特的资源，在真实教育环境中为鲁棒性活动识别模型的开发提供了丰富的多模态数据。

## 引用方式
如果本项目对你的研究有所帮助，请使用以下BibTeX条目引用我们的工作：
@misc{marquezcarpintero2025caddiinclassactivitydetection,
title={CADDI: An in-Class Activity Detection Dataset using IMU data from low-cost sensors},
author={Luis Marquez-Carpintero and Sergio Suescun-Ferrandiz and Monica Pina-Navarro and Miguel Cazorla and Francisco Gomez-Donoso},
year={2025},
eprint={2503.02853},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2503.02853},
}

搜集汇总

背景与挑战

背景概述

CADDI数据集是一个专注于课堂活动识别的多模态数据集，包含12名参与者完成的19种活动数据，使用Samsung Galaxy Watch 5的IMU传感器和立体相机图像采集，支持传感器融合技术的研究。数据集包含连续和瞬时活动，时间跨度为472.03分钟，IMU数据采样率为100Hz，相机图像为25-30 fps，适用于教育环境中的活动识别模型开发。