Hand Gesture Recognition in Video with Single Short Detector

With the continuous growth of the global population, innovative human-computer interaction (HCI) technologies are becoming increasingly important in improving quality of life. Among these, gesture-based systems stand out for their ability to enhance accessibility, safety, and user experience, particularly for individuals with physical impairments, while also benefiting the wider community. However, recognizing gestures from video data remains a complex challenge due to variations in motion patterns across different users.This research makes use of the Hand Gesture Classification dataset to examine and compare multiple algorithms for gesture recognition. For the classification task, several approaches were implemented, including Convolutional Neural Networks (CNN) integrated with Support Vector Machines (SVM), Deep Belief Networks (DBN) with SVM, Histogram of Oriented Gradients (HOG) with SVM, Histogram of Optical Flow (HOF) with SVM, as well as an ensemble strategy that combines Xception, CNN, SVM, and a Voting Classifier (utilizing Boosted Decision Trees and Random Forest). For gesture detection, the YOLO family of models was applied, specifically YOLOv5x6, YOLOv5s6, YOLOv8n, and YOLOv9n.The experimental findings reveal that the Xception-based CNN ensemble delivers the highest accuracy, outperforming other models and proving to be particularly effective in reliable gesture recognition from video sequences.