Experimantal Study on Arm Part Detection for Pose Estimation
Pose estimation is very important in fields like forensic research and surveillance applications. Control by gesture and pose are increasingly used in human-computer interaction and since the introduction of Microsoft’s Xbox 360 Kinect device (1) it is becoming a standard in the gaming industry. High accuracy of pose estimation can be achieved through the use of high quality stereo camera’s, but of course there is still room for improvement. However, good pose estimation using low resolution devices like webcams is still a very difficult task. Vincent van Megen (2) presented a system capable of estimating upper body pose in low resolution images with reasonable accuracy. In this thesis, local classifiers are explored which are trained to improve such a rough estimation by localizing the individual parts of human limbs. The use of different image features are compared as well as the use of different classification algorithms. The different detection and localization methods are evaluated on two different data sets, testing the generalizability of the detectors. HoG features and covariance features combined with neural network based arm part detectors turn out to be the most accurate, fast, and best generalizing on the tested data sets.
Faculteit der Sociale Wetenschappen