Dual-Location Smartphone IMU Human Activity Recognition Dataset (Hand & Ankle, 60s Trials)

This dataset was created to study smartphone-based Human Activity Recognition (HAR) using raw inertial signals recorded at two common body locations (hand and ankle). The working hypothesis is that longer, continuous recordings (approximately 60 seconds per activity) and multi-sensor fusion (accelerometer, gyroscope, magnetometer) provide richer temporal and biomechanical signatures that improve activity classification, and that non-linear ensemble models capture these relationships more effectively than linear classifiers.
What the data contains
• Raw sensor streams exported as CSV files from a smartphone IMU logger (Phyphox).
• Each trial typically includes three separate CSV files: accelerometer, gyroscope, and magnetometer.
• Each CSV contains a time column (seconds) and tri-axial measurements (X, Y, Z) with sensor-specific units.
• Two placement conditions are included: hand-held and ankle/foot-mounted.
• Ground-truth labels correspond to six daily activities; labels may be encoded jointly with placement (e.g., Walking_Hand, Walking_Foot) to form a 12-class setup.
How the data was gathered
• Participants performed each activity for approximately 60 seconds per placement.
• Recordings were made in real time using a standard smartphone and the Phyphox application, then exported to CSV.
• Participants were anonymized; the study followed a minimal-risk, non-invasive protocol with informed consent.
How to interpret and use the data
• Use the time column as the temporal index for each sensor stream.
• For multi-sensor modeling, align streams by time and merge accelerometer, gyroscope, and magnetometer to form a 9-channel signal (Acc/Gyro/Mag × X/Y/Z).
• For window-based learning, segment continuous signals with overlapping sliding windows (e.g., 150 samples per window, stride 100 samples).
• If using classical ML, compute window-level statistical features (mean, standard deviation, min, max, range, and signal energy) for each channel; this yields 54 features per placement (9 channels × 6 statistics), or 108 features when concatenating hand and ankle features (feature-level fusion).