A real-time recognition gait framework for personal authentication via image-based neural network: accelerated by feature reduction in time and frequency domains

In recent years, personal authentication based on attitude estimation—gait recognition authentication has become a popular research topic because of its long-range, non-invasive, non-contact, high-precision, and other advantages. However, at present, most relevant research prefers to use the acquired original data directly for iteration and learning. As a result, it takes too long to learn relevant models in the use scenarios with complicated data and heavy human traffic, such as airports and railway stations, where real-time identification cannot be completed while maintaining accuracy, and thus a scheme to improve the learning and recognition speed is needed. Therefore, in this paper, we proposed an innovative real-time MediaPipe-based gait analysis framework and a new Composite Filter Feature Selection (CFFS) method via key nodes, angles, and lengths calculating. Then, based on the proposed method, we extract the aimed features as a new dataset and verified it by 1D-CNN neural network. Furthermore, we also applied Hilbert–Huang transform to investigate these extracted gait features in the frequency domain, improving the performance of our proposed framework to achieve real time under higher recognition accuracy. The experimental results show that the innovative gait recognition framework and data processing technology can reduce the gait feature data, speed up the process of gait recognition, and still maintain the original recognition accuracy. It can also be applied to various large, enclosed spaces with the huge human flow, which has played a role in improving the safety factor, saving labor costs, and accelerating economic consumption.