Comparative study of static hand gesture classification methods based on hand landmark coordinates

Problem statement. Real-time hand gesture interfaces require accurate classification with low latency and without specialized hardware. Instead of processing full video frames, a gesture can be represented by the coordinates of hand landmarks detected by MediaPipe Hands. This greatly reduces the size of the classifier input, but it remains necessary to determine which classification method is most suitable for such data.

Goal. To compare classical machine-learning algorithms and neural-network models for static hand gesture recognition based on MediaPipe Hands landmark coordinates and to select a model suitable for an interactive application.

Results. The dataset contained 510 observations from six gesture classes, with 85 samples in each class. Random Forest, linear and RBF Support Vector Machine, K-Nearest Neighbors with \(k=3\) and \(k=5\), a multilayer perceptron, and a one-dimensional convolutional neural network were evaluated. The first two principal components explained 94.9% of the variance. Four models achieved Test F1 = 1.0. The lowest classification time, 0.38 ms, was measured for the multilayer perceptron, although its Test F1 was lower, at 0.922. Among the models with Test F1 = 1.0, linear SVM was the fastest, with an inference time of 0.41 ms.

Practical significance. The experiment shows that a relatively simple classifier can be sufficient when the input is already represented by hand landmark coordinates. Linear SVM combined perfect classification of the test subset with low inference time and was therefore selected for the prototype. The reported timing refers only to the classification stage and does not include video capture, MediaPipe landmark detection, or interface rendering. Since all observations were collected from one participant under controlled conditions, the model should be tested separately on data from other users and under different recording conditions.