<![CDATA[Human-Computer Interaction (HCI) is increasingly shifting toward touchless and intuitive interfaces, where simple navigational gesture recognition such as left, right, and stop constitutes a fundamental element. Gesture recognition methods based on Histogram of Oriented Gradients (HOG) and 9-Uniform Local Binary Pattern (9ULBP) features with Support Vector Machine (SVM) have proven effective for hand gestures, yet they remain highly vulnerable to complex background clutter. Applying these methods to upper limb gestures (involving arms and upper body) faces two major challenges: first, hand‑only segmentation is inadequate for capturing arm orientation information; second, there is no systematic evaluation of automatic segmentation methods based on depth versus landmarks to improve HOG feature robustness in the upper‑limb gesture domain. This study proposes integrating two preprocessing frameworks—MiDaS (monocular depth estimation) and MediaPipe Holistic (spatial guidance based on pose and hand landmarks)—prior to HOG and HOG+9ULBP feature extraction and SVM classification. Evaluation on 1,400 training images and 74 test images shows that the MediaPipe Holistic + HOG pipeline achieves the best performance with 82.4% accuracy and a macro F1‑score of 0.833, substantially outperforming MiDaS + HOG (60.8% accuracy; F1=0.473). Adding 9ULBP features decreases the F1‑score of the best pipeline to 0.715, proving that local texture is irrelevant for upper‑limb gestures. Consequently, landmark‑based human segmentation is recommended over depth‑based approaches, and feature extraction using only HOG (without 9ULBP) is sufficient.]]>
Copyrights © 2026
