PSU Multiview Profile-based Action RBGD dataset
Introduction
Recognizing profiled-based actions of human is a challenging thing in computer vision that lead to understanding complex actions, activities and behaviors especially for video surveillance and health-care applications. In this dataset we introduce a profile-based human action vision data from couple-view using RGBD cameras. The dataset is captured simultaneously that synchronized with frame sequence in format of videos. This dataset include 4 profile-based actions consist of standing / walking, sitting, bending and laying.
We created two different scenarios: clear room and living room scenario.
 |
 |
|
Example scenario of clear room |
Example scenario of living room |
Scenario Setup
We setup Kinect cameras both scenarios at overhead-view using poles 2 meter with angle 60° with pole.
 |
|
Camera Setup in Side view |
We capture RGB and Depth information from multi-views taken from two cameras with varies orientation of human at different static-viewpoints in the overlapping Area of Interest. The operation range is about 3 to 5.5 meters from cameras that could acquire the full human body. The dataset used CLNUI library for acquiring the depth information from Kinect cameras.
The scenario of clear room includes 8700 action frames, whereas the angle between cameras are perpendicular and range is about 3-4m. We fixed orientation that point to human 5 types: Front, Slant, Side, Back-slant and Back.
The scenario of living room include 7800 action frames, with four different angles: 30°, 45°, 60° and 90° degrees and range is about 4-5m. Which human can be free turn around every orientation and move in area of interest.
 |
 |
 |
|
Clear Room Setup and orientation in Top view. |
Living Room Setup in Top view. |
|
The video dataset was captured at size of 640×480 for RGB, DEPTH 8 bits and DEPTH 24 bits (represent as color) that provided by CLNUI library. In the each sequence is performed by 3-5 actors. The video will contain background at least 40 frames at beginning of video. The background is provided will be used motion detection of background subtraction algorithm for separate human from background. The real frame rate was captured from library is about 8-12 fps. In the each set of video will be contain in directory as name of set e.g. Sitting90, bending_back ,etc. In the each directory will include all views of RGB, DEPTH 8 bits and DEPTH 24 bits, like as “1colorRGB.avi, 1depthGRAY.avi, 1depthRGB.avi, 2colorRGB.avi, 2depthGRAY.avi, 2depthRGB.avi”. The video is recorded as XVID codec using OpenCV.
Video sequence in the each set
Download
Please read and accept Agreement and Disclaimer below before download. If you download this dataset, we count you accepted agreement and disclaimer.
1. The dataset must be only used for non-commercial or educational purposes.
2. No responsibility for any incidents, or damages caused by the direct or indirect usage this dataset.
Here is example of video dataset.
— Scenario of clear room: Download Example Video RGB / Depth8 / Depth24 2colorRGB.avi / 2depthGRAY.avi / 2depthRGB.avi
— Scenario of living room: Download Example Video RGB / Depth8 / Depth24 1colorRGB.avi / 1depthGRAY.avi / 1depthRGB.avi
Here is all of video dataset.
— Scenario of clear room: Download Here (7zip 2.70 GB) Clear_Room_Scenario.7z
— Scenario of living room: Download Here (7zip 1.62 GB) Living_Room_Scenario.7z
Sample Frame
