Towards Identifying High-Performing Feature Descriptors for Depth-Based Hand Signs Recognition

This paper presents an innovative strategy for improving hand sign recognition by utilizing some robust handcrafted features derived from Local Binary Patterns (LBP), Histogram of Oriented Gradients (HOG), Scale-Invariant Feature Transform (SIFT), Multi-Radii Circular Signatures (MRCS), and Multi-Scale Density (MSD). The datasets are drawn from diverse cultural backgrounds and sources such as Creative Senz3D, HKU EEE DSP, and NTU 10-gesture and they include a wide array of hand gestures. LBP and HOG features adeptly capture intricate texture and shape details, while SIFT descriptors ensure resilience to variations in size, rotation, and illumination. Furthermore, MRCS strategically employs multiple circular signatures across the image object, and MSD utilizes a hierarchical approach to segment the input image into smaller zones. Through comprehensive experimental evaluations employing various classifiers, our results underscore the efficacy of this novel technique. Of the variety of classifiers used, MLP has proven to be particularly effective in the MSD scenario, outperforming previous approaches in terms of precision, recall, F-score, and accuracy (0.9815, 0.9978, 0.9834, and 99.78%, respectively) on the Creative Senz3D dataset.