Performance Comparison and Analysis of Gesture Recognition Based on YOLOv5 and Traditional Machine Learning

With the proliferation of electronic devices, traditional interaction methods such as the mouse and keyboard have gradually shown their limitations, while gesture recognition has gained increasing attention due to its convenience. This study compared the advantages and disadvantages of different feature extraction methods, using Fourier Descriptors (FFT), Histogram of Oriented Gradients (HOG), and yolov5 for feature extraction, and evaluated their noise resistance capabilities. In terms of machine learning and real-time processing, the three-frame difference method was used, which can recognize small gesture amplitudes and effectively filter out static skin-like regions. The results showed that HOG was more accurate than FFT, but slower when processing large images and very sensitive to object size, which is detrimental to SVM classification. yolov5 performed better in real-time applications, but its noise resistance was insufficient in complex backgrounds, indicating that further model adjustments are needed to adapt to noisy environments.