Source-linked AI summary
HuGaDB: Human Gait Database for Activity Recognition from Wearable Inertial Sensor Networks
Roman Chereshnev, Attila Kertesz-Farkas
TL;DR
HuGaDB addresses the need for detailed gait data that can support both activity recognition and analysis of how people move. It collects continuous, annotated activity sequences with wearable inertial and EMG sensors, providing data for gait analysis and related applications.
Problem
Existing wearable-sensor databases vary in their sensor types and activities, while gait analysis also needs data for studying how activities are performed.
Source: §1 · pp. 1–2Method
HuGaDB records continuous combinations of annotated activities using six inertial sensors on the thighs, shins, and feet and two EMG sensors on the quadriceps.
Source: §2 · p. 2Results
Takeaways & Limitations
The dataset can support gait and activity analysis, including health-care studies and modeling human movement in virtual reality or humanoid robotics.
Source: §2 · p. 2, §7 · p. 5Data were collected mainly indoors, and the dataset contains no treadmill recordings.
Source: §3 · p. 3
Abstract
from arXiv · showhide
This paper presents a human gait data collection for analysis and activity recognition consisting of continues recordings of combined activities, such as walking, running, taking stairs up and down, sitting down, and so on; and the data recorded are segmented and annotated. Data were collected from a body sensor network consisting of six wearable inertial sensors (accelerometer and gyroscope) located on the right and left thighs, shins, and feet. Additionally, two electromyography sensors were used on the quadriceps (front thigh) to measure muscle activity. This database can be used not only for activity recognition but also for studying how activities are performed and how the parts of the legs move relative to each other. Therefore, the data can be used (a) to perform health-care-related studies, such as in walking rehabilitation or Parkinson's disease recognition, (b) in virtual reality and gaming for simulating humanoid motion, or (c) for humanoid robotics to model humanoid walking. This dataset is the first of its kind which provides data about human gait in great detail. The database is available free of charge https://github.com/romanchereshnev/HuGaDB.
1 Introduction
Wearable-sensor databases support gesture, daily-activity, and gait research, but vary in sensors and recorded activities. Gait analysis additionally examines how activities are performed, supporting health-care applications.
1 Introduction
- Wearable body sensors are enabling new scientific studies and industrial applications across gesture recognition, human activity recognition, and human gait analysis.The authors frame these as large application areas for increasingly available wearable sensors.Source: §1 · p. 1
- Existing wearable-sensor databases span gesture recognition, daily-lifestyle activity recognition, and walking-related activity recognition, with differing sensor types and activities.Examples include hand-gesture datasets, daily-routine and kitchen datasets, and walking-related datasets such as WARD, PAMAP2, USC-HAD, and MAREA.Source: §1 · p. 1
- Benchmarking efforts include shared datasets and live evaluations that assess recognition accuracy alongside practical usability criteria.EvAAL-AR evaluates user acceptance, recognition delay, installation complexity, and interoperability as well as accuracy.Source: §1 · pp. 1–2
- Gait analysis studies both recognized activities and how they are performed, including applications such as patient monitoring, fall detection, and Parkinson’s disease diagnosis.The cited examples include a Parkinson’s dataset with recordings from 10 affected participants and a separate gait reference dataset from 40 healthy people.Source: §1 · p. 2
2 Motivation and Design Goals
HuGaDB is designed to capture detailed leg movement across continuous, annotated sequences of activities, including transitions. Its inertial-sensor recordings are intended to support gait research and applications from health care to robotics.
2 Motivation and Design Goals
- Continuous recordings combine activities and label the activity currently being performed, enabling analysis of movement within activities and between transitions.The stated purpose includes studying how leg parts move individually and relative to one another during activities such as walking, running, and standing up.Source: §2 · p. 2
- The authors favor inertial sensors because they are inexpensive, simple to use indoors or outdoors, and widely available.They contrast these advantages with video-based motion capture’s costly cameras and suits, fixed test areas, lighting sensitivity, and lost-marker problem.Source: §2 · p. 2
- Six inertial sensors on both thighs, shins, and feet recorded 10 hours from 18 healthy participants for studying individual and relative leg movement.The authors propose possible uses as control data in rehabilitation or Parkinson’s disease studies, and for modeling movement in virtual reality, gaming, or humanoid robotics.Source: §2 · p. 2
- HuGaDB provides detailed gait recordings with segmented annotations that support studying transitions between activities.The dataset is described as the first to provide human gait data in great detail, mainly from inertial sensors, with annotations for transitions.Source: §2 · p. 2
3 Data Collection and Sensor Network Topology
HuGaDB uses a bilateral network of inertial and EMG sensors to record continuous activity sequences from healthy participants. The collection totals 10 hours, with indoor data predominating and no treadmill recordings.
3 Data Collection and Sensor Network Topology
- The inertial sensors combine a three-axis accelerometer and gyroscope, while EMG channels sample at 1.0 kHz and store 8-bit data.The inertial sensors use ±2g accelerometer and ±2000°/s gyroscope ranges; data were retained raw for alternative normalization.Source: §3 · pp. 2–3
- Three pairs of inertial sensors were placed symmetrically on the thighs, shins, and feet, alongside EMG sensors on the vastus lateralis, yielding 38 signals.Each inertial sensor records three-axis acceleration and gyroscope signals; the two EMG sensors use three electrodes each.Source: §3 · p. 3
- Participants were 18 healthy young adults: 4 women and 14 men, with a mean age of 23.67 years.Their average height was 179.06 cm and average weight was 73.44 kg.Source: §3 · p. 3
- The collection comprises 2,111,962 samples and 10 hours of continuously recorded, activity-segmented data from 18 participants performing combinations of activities.Participants performed activities at normal speed and casually, without obstacles; an experimenter annotated the activity labels during recording.Source: §3 · p. 3
- Data collection occurred mainly indoors, with some activities performed outdoors and in moving vehicles or an elevator; no treadmill data were collected.The authors note that force impacts on accelerometers in moving settings may matter in some applications.Source: §3 · p. 3
4 Data Format
HuGaDB stores recordings in flat, tab-delimited text files with metadata headers and a fixed sensor-column layout. The format includes activity labels and scripts for loading or importing data.
4 Data Format
- A file represents either one activity or a series of activities, and filenames encode the activity, participant, and recording counter.Files containing multiple activity types use VARIOUS; for example, HGD v1 walking 17 02.txt identifies participant 17’s second walking recording.Source: §4 · pp. 3–4
- Files contain tab-delimited raw sensor data beneath a header summarizing activities, activity IDs, and recording time and date.The header and file-naming convention are described in Tables 3 and 2, respectively.Source: §4 · pp. 3–4
- Each data-file row is a sample, and its 39 fixed columns contain 36 inertial signals, 2 EMG signals, and an activity ID.The inertial columns are ordered by sensor location and each sensor contributes three acceleration and three gyroscope axes.Source: §4 · p. 4
- The paper illustrates sample data across activities in a heat-map visualization and describes normalization for inertial and EMG signals.The inertial values are divided by 32768; EMG values are shifted by subtracting 128 and then divided by 128.Source: §4 · p. 4, Figure 2 · p. 4
- The authors provide loading examples for Python and Matlab and a script for loading files into SQLite.The Python example requires NumPy.Source: §4 · p. 5
5 Discussion on Data Variance
Walking data show greater frame-level variance across users than within repeated recordings by one user, motivating evaluation that holds out users. Sensor-brand differences may also contribute variance, but were not measured.
5 Discussion on Data Variance
- Different users show much higher same-frame variance than repeated recordings from one user, potentially reflecting gait, leg shape, or sensor placement differences.EMG data showed even higher variance, which the authors associate with differences in skin electrical conduction and thickness.Source: §5 · p. 5
- Holding out one participant’s data for testing estimates recognition performance on a new user whose data were not used for training.The authors train on the other 17 participants and propose supervised cross-validation because of high between-user variance.Source: §5 · p. 5
- The study did not collect data from different sensor brands, leaving brand-related variance unmeasured and sensor compatibility dependent on calibration.The authors hope measurement noise is generally small and different sensors can be calibrated to be compatible.Source: §5 · p. 5
7 Summary
HuGaDB provides detailed, segmented human-gait recordings from 18 participants, totaling around 10 hours. The authors identify health-care, virtual-reality, and humanoid-robotics applications.
7 Summary
- HuGaDB combines detailed inertial-sensor gait data with segmented annotations for transitions between activities.The dataset is intended for human gait analysis and activity recognition.Source: §7 · p. 5
- The dataset comprises recordings from 18 participants, totaling around 10 hours.Source: §7 · p. 5
- The authors identify walking rehabilitation and modeling human movement in virtual reality or humanoid robotics as potential uses.They state that the dataset will be updated with data from new participants.Source: §7 · p. 5