Efficient features for representing hand shape in images : By using linear projections in the HOG feature space
Year: 2014
Type: dissertation
Abstract: This thesis explores hand pose estimation, which means mapping a 2D imageto a hand pose. Hand pose estimation has many promising applications suchas hand sign recognition, robotic learning by demonstration, and human computer interaction in general. To do the estimation, image features are extracted from the image and a mapping to the space of hand poses is the nconstructed. Ideally the mapping from image features to pose space would be one-to-one, but in reality it is rather a many-to-many mapping leadingto ambiguities. This is due to the image feature not capturing the actualpose, but the form of the 2D projection of the hand. Hands may also occludeparts of itself which also leads to ambiguities. This thesis explores ways in which to improve the commonly used image feature HOG (Histogram of Oriented Gradients), by capturing the HOG subspace used by hand images, to obtain a feature whose mapping to pose space is more well-behaved than that of the HOG feature. The new feature is computed as projection on anddistances to lines in HOG space. The new feature’s performance is tested against the HOG feature using nearest neighbour (NN) regression and the results show that the new feature does not yet perform as well as the HOG feature. Nevertheless, the conclusion is that the new feature, called MPDD,for Multiple Projection and Distance Dimensions, does indeed capture the most relevant information in HOG, but fail to use it as well as the HOG does with the current construction method. However, constructing the MPDD in a slightly different way could potentially lead to improvements and so future research could still be of interest.
Author: Patrik Berggren
Language: English
Subfield: Human-Computer Interaction
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W1605033482
