Environment and Data Acquisition Description

For the development of the GRIDDS dataset we had the collaboration of 35 volunteers, among students, teachers and staff from the Polytechnic Institute of Viana do Castelo (IPVC). The recording sessions occurred in June of 2018, at the IPVC facilities, in a controlled indoor environment, with a static background and with both natural and artificial lighting.

Two trajectories were defined, in a straight line across the room: one starting from the left side of the room to the right side, and the other on the opposite direction. The Kinect v2 sensor, supported on a tripod, was fixed at 1.8 meters high, perpendicular to the defined trajectories. Each one of the 35 volunteers completed 5 walking sequences per trajectory, at a self-selected comfortable pace, resulting in a total of 10 sequences per participant.

Data Specification and Availability

The dataset is composed by 35 folders (one per participant), each one containing the following collected data: a set of color, depth and infrared images, a set of depth data files, a file with the joints' coordinates and and a file with the corresponding timestamps of each captured frame from the previously mentioned streams. Additionally, we included body silhouette images, cropped, all facing to the same side, and normalized in size, with a resolution of 80x120 pixels. Inside of each folder, the available information is in either one of the following formats:

- vvv_ss_stream_nnn.fmt, for the color, depth, silhouette and infrared streams;

- vvv_ss_stream.fmt, for the timestamp and joints' coordinates streams;

where vvv corresponds to the volunteers' id, ss to the session number, stream to the different available streams, nnn to the frame number and fmt to the different file formats (PNG or CSV).

For example, the file named 003_09_depth_021.csv corresponds to the frame number 21 of the depth stream of the volunteer with id equals to 003, captured during the session number 09, saved in the CSV file format.

All image files are in the Portable Network Graphics (PNG) format, varying only in the bit-depth color information: while the color images are in 24-bit, the depth images, which are in gray-scale, are in 16-bit, the body silhouettes are in 1-bit and the infrared images are in 16-bit.

The depth data files (which are in the Comma-Separated Values (CSV) format) have the same resolution as the depth images, however, in this case, each cell contains a value corresponding to the distance (in millimeters) between the Kinect device and the object(s) detected in front of the device.

The coordinate files are also in the CSV format and have a resolution of 7 columns by (N_frames x N_joints) lines, where each one of the 7 columns corresponds to: the frame number; the 3D coordinates (x,y,z) and the 2D coordinates (x,y) of the tracked joint (both in meters); and finally, the last column corresponds to the tracking state of the corresponding joint (1 means that the joint data was inferred and confidence in the position data is lower than if it were Tracked; 0 means that the joint data was not tracked and no data is known about this joint; 2 means that the joint data was tracked and the data can be trusted). The number of lines is based on the number of captured frames (N_frames) multiplied by the number of tracked joints (25 joints) (N_joints)