Low-complexity supervised learning for gesture and shape recognition
Year: 2014
Type: dissertation
Abstract: Classi cation is a machine learning task in which the objective is to categorize given samples according to their attributes. Gesture Recognition (GR) and Shape Recognition (SR) are two classi cation examples. Some daily-life applications of these include Hand Gesture Recognition (HGR) and Optical Character Recognition (OCR). GR is a challenging classi cation problem often used in human-computer interaction applications to provide a natural interface between user and computer. Since the same gesture might be performed with di erent speeds, Dynamic Time Warping (DTW) is needed to nd the optimal alignment between two time sequences. Oftentimes a preprocessing of sequences is required to remove variations between the reference gestures and the test gestures. We discuss a set of pre-processing methods to make the gesture recognition mechanism robust to these variations. DTW computes a dissimilarity measure by time-warping the sequences on a per sample basis by using the distance between the current reference and test sequences. However, all body joints involved in a gesture are not equally important in computing the distance between two sequence samples. We propose a weighted DTW method that weights joints by optimizing a discriminant ratio. SR is another classi cation problem with increasing number of applications from OCR to pedestrian detection. Decision tree is a good choice of classi er for shape recognition because it is easy to implement and visualize and has lower computational complexity. Bagging randomized decision trees as random forests increases the accuracy rates if the trees are weakly correlated. We propose using random rectangles in combination with random forests and test our method on OCR and GR datasets. We show that the accuracy of our method is similar to the OCR state-of-the-art and better than the GR state-of-the-art, while executing signi cantly faster, which makes our proposed method a good t for real-time object/shape recognition. Then discuss how a simple feature such as a random rectangle can perform similar to the complex statistical and structural features designed for shape recognition. Finally we analyze the e ect of our parameters.
Author: Çelebi, Sait
Language: English
Subfield: Human-Computer Interaction
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W2281523041