Quantitative Assessment of Hand Signal Recognition Using Landmarks Detection: A Comparative Study of Machine Learning Techniques

Abstract

Hand movement comprehension and interpretation are critical in various domains, including robotics, human-computer interface, and sign language identification. Computer vision techniques have lately been utilized to improve the accuracy and usefulness of hand signal identification. This study investigates alternative machine learning algorithms for identifying hand gestures that involve landmark detection. In computer vision, it is common to identify specific areas on an object in order to extract relevant properties from it, in this case, the hand. This procedure is known as “landmark detection.” In this paper, researchers use the Python OpenCV package to extract landmarks from images of hand gestures, which research then uses as input features for machine learning models. The performance of four machine learning algorithms is compared: Support Vector Machine (S-VM), Random-Forest (RF), k-Nearest-Neighbors (K-NN), and Multi-Layer Perceptron (MLP). The research evaluates each model's performance using a unique dataset of hand signal images obtained with a Kinect sensor. With an average accuracy of 95%, the data show that the S-VM technique surpasses the other three algorithms. The RF algorithm performs admirably, with an average accuracy of 92%. The MLP algorithm performs the worst, with an average accuracy of 83%, followed by the K-NN method, which performs moderately, with an average accuracy of 88%. Overall, the study emphasizes the need to select the appropriate machine learning algorithms and the importance of landmark detection for identifying hand gestures. While K-NN and MLP algorithms may require further feature engineering or optimization, S-VM and RF approaches are acceptable for hand signal recognition utilizing landmark detection.