Unencrypted login | home

Program Information

Accuracy Evaluation of Depth Data in Microsoft Kinect

no image available
K Kozono

K Kozono1*, M Aoki1, M Ono1, Y Kamikawa2, H Arimura3, F Toyofuku3, (1) Department of Health Sciences, Graduate School of Medical Sciences, Kyushu University, (2) Department of Health Sciences, School of Medicine, Kyushu University, (3) Department of Health Sciences, Faculty of Medical Sciences, Kyushu University,

SU-E-I-92 Sunday 3:00:00 PM - 6:00:00 PM Room: Exhibit Hall

Purpose: Microsoft Kinect has potential for use in real-time patient position monitoring in diagnostic radiology and radiotherapy. We evaluated the accuracy of depth image data and the device-to-device variation in various conditions simulating clinical applications in a hospital.

Methods: Kinect sensor consists of infrared-ray depth camera and RGB camera. We developed a computer program using OpenNI and OpenCV for measuring quantitative distance data. The program displays depth image obtained from Kinect sensor on the screen, and the cartesian coordinates at an arbitrary point selected by mouse-clicking can be measured. A rectangular box without luster (300 x 198 x 50 mm³) was used as a measuring object. The object was placed on the floor at various distances ranging from 0 to 400 cm in increments of 10 cm from the sensor, and depth data were measured for 10 points on the planar surface of the box. The measured distance data were calibrated by using the least square method. The device-to-device variations were evaluated using five Kinect sensors.

Results: There was almost linear relationship between true and measured values. Kinect sensor was unable to measure at a distance of less than 50 cm from the sensor. It was found that distance data calibration was necessary for each sensor. The device-to-device variation error for five Kinect sensors was within 0.46% at the distance range from 50 cm to 2 m from the sensor. The maximum deviation of the distance data after calibration was 1.1 mm at a distance from 50 to 150 cm. The overall average error of five Kinect sensors was 0.18 mm at a distance range of 50 to 150 cm.

Conclusions: Kinect sensor has distance accuracy of about 1 mm if each device is properly calibrated. This sensor will be useable for positioning of patients in diagnostic radiology and radiotherapy.

Contact Email