A multi-sensor gait dataset collected under non-standardized dual-task conditions

Non-standardized dual-tasks have recently gained attention in health monitoring and post-operative rehabilitation. By collecting data with multiple sensors, we can quantify motion characteristics from different perspectives and explore the complementarity and interchangeability between sensors. Currently, there is a lack of publicly available non-standardized dual-task gait datasets collected with multiple sensors, thus we proposed a dataset (NONSD-Gait) from 23 healthy adults walking back and forth over 7 meters under three dual-task conditions, collected by three types of sensors: optical motion capture (MOCAP) system, depth camera and inertial measurement unit (IMU). MoCap captured the 3D trajectories of 22 markers attached to the subject using 8 optical cameras, while the depth camera recorded the 3D trajectories of 25 joints through a non-contact depth camera. The IMU was placed on the left ankle to record 3-axis acceleration and angular velocity data. Each participant underwent two repeated experiments for each task. Moreover, this dataset also includes extracted spatio-temporal gait parameters and kinematic parameters, supporting gait feature recognition in complex scenarios and multimodal gait data analysis.

Methods
This study recruited 23 healthy adults aged between 21 and 30 years old (9 males, 14 females). All participants had no neuromuscular diseases or skeletal injuries, and without any hearing or vision impairment. In the experiment, participants were required to walk back and forth on a 7m × 1m mat under a single-task condition and three non-standardized dual-task conditions. This study utilized eight MOCAP cameras (NOKOV MARS 2H HD cameras, 100Hz, resolution 2048 × 1088), one depth camera (Microsoft Kinect V2.0, 30Hz, resolution 512 × 424), and one inertial sensor (Witmotion BWT901BLECL5.0C, 100Hz) for data collection.

For the walking data from each task recorded by MOCAP, 22 reflective markers  were firstly labeled using  MOCAP's software and the 3D trajectories of the markers were then exported. The data collected by IMU were exported using IMU's software, and the data collected by Kinect were exported using its proprietary SDK. The data collected by the three sensors were all exported in .csv format. After that, for partially missing values of certain markers  in the three-dimensional trajectories exported by MOCAP, a relational interpolation method based on the movement of surrounding markers was applied using Anaconda3. To remove high-frequency noise, the data collected by MOCAP and IMU were smoothed using a Butterworth low-pass filter with a cutoff frequency of 7.5 Hz. The noise and fluctuations of the data collected by Kinect were reduced using a moving average with a window size of 3. All the data were then segmented using six manually recorded timestamps from the experiment. To minimize the impact of gait initiation on walking speed, the final gait dataset excluded the initial and final 2 m of walking. The data collected by MOCAP and IMU were segmented into 5m back-and-forth data, 2 m back-and-forth   data, and turning data, while the data collected by Kinect was segmented into 2 m back-and-forth data and turning data.  Attribute to the Kinect's erroneous identification of left and right joints as reversed during the backward phase, we adjusted the coordinates by swapping the left and right joints during this phase. After data preprocessing, we extracted a total of nine spatio-temporal gait parameters and fourteen kinematic parameters.