Degree
Master of Engineering Science
Program
Electrical and Computer Engineering
Supervisor
Jagath Samarabandu
Abstract
Conventional human action recognition methods use a single light camera to extract all the necessary information needed to perform the recognition. However, the use of a single light camera poses limitations which can not be addressed without a hardware change. In this thesis, we propose a novel approach to the multi camera setup. Our approach utilizes the skeletal pose estimation capabilities of the Microsoft Kinect camera, and uses this estimated pose on the image of the non-depth camera. The approach aims at improving performance of image analysis of multiple camera, which would not be as easy in a typical multiple camera setup. The depth information sharing between the camera is in the form of pose projection, which depends on location awareness between them, where the locations can be found using chessboard pattern calibration techniques. Due to the limitations of pattern calibration, we propose a novel calibration refinement approach to increase the detection distance, and simplify the long calibration process. The two tests performed demonstrate that the pose projection process performs with good accuracy with a successful calibration and good Kinect pose estimation, however not so with a failed one. Three tests were performed to determine the calibration performance. Distance calculations were prone to error with a mean accuracy of 96% under 60cm difference, and dropping drastically beyond that, and a stable orientation calculation with mean accuracy of 97%. Last test also proves that our new refinement approach improves the outcome of the projection significantly with a failed pattern calibration, and allows for almost double the camera difference detection of about 120cm. While the orientation mean calculation accuracy achieved similar results to pattern calibration, the distance was less so at around 92%, however, it did maintain a stable standard deviation, while the pattern calibration increased as distance increased.
Recommended Citation
Kayal, Omar Ghassan, "Use of Microsoft Kinect in a dual camera setup for action recognition applications" (2014). Electronic Thesis and Dissertation Repository. 2072.
https://ir.lib.uwo.ca/etd/2072