DaRA Dataset: Combining Wearable Sensors, Location Tracking, and Video for Daily Living Activity Recognition

Overview

The DaRA Dataset (Daily living activity Recognition with wearables and video) published in early 2026 provides a rich multimodal collection for human activity recognition research, combining inertial sensors, indoor positioning, and visual data.

Data Collection Setup

• 18 subjects recruited with diverse demographics performing scripted activities in laboratory environments designed to mimic real-world home and office settings.
• Each subject wore two sets of three MotionMiners devices (wearable IMUs) on both wrists and torso for redundancy and data quality analysis.
• Bluetooth Low Energy (BLE) beacons deployed throughout the environment for indoor localization via RSSI (Received Signal Strength Indicator).
• First-person (egocentric) video cameras worn by participants to capture visual context and enable multi-modal annotation.

Sensor Specifications

• IMU sensors: Tri-axial accelerometer and gyroscope at 100 Hz sampling rate from three body locations (both wrists, torso).
• BLE localization: RSSI measurements at 10 Hz from multiple beacon emitters for room-level and sub-room indoor positioning.
• Video data: First-person perspective video synchronized with sensor data for ground truth validation and context.
• Dual sensor sets: Redundant IMUs enabling robustness analysis and sensor failure mitigation research.

Activity Categories (12 Classes)

• Locomotion: Walking, standing still, gait cycles, step patterns.
• Transitions: Sit-to-stand, stand-to-sit, lying down, getting up.
• Upper body activities: Reaching, grasping, bending, manipulating objects.
• Complex activities: Dual-task scenarios, carrying objects, household tasks.
• Fall-related events: Near-falls, balance perturbations (supporting fall detection research).

Multi-Level Labeling

• Single-label categories with exactly one activity label per time interval (e.g., walking, standing).
• Multi-label categories allowing concurrent activity labels (e.g., walking while carrying, bending while reaching).
• Hard labels with unambiguous 100% allocation for precise ground truth.
• Temporal segmentation enabling both frame-level and episode-level analysis.

Use Cases

• Multimodal activity recognition: Fusing IMU, location, and video data for robust HAR systems.
• Indoor localization research: Developing BLE-based positioning algorithms integrated with activity context.
• Sensor placement analysis: Comparing wrist vs. torso IMU performance for different activity types.
• Fall detection: Training algorithms to detect pre-fall patterns and actual falls using wearable sensors.
• Wearable device evaluation: Benchmarking commercial and research-grade IMU devices in realistic scenarios.