<p>Photoplethysmography (PPG) signals were shown to have potential in human motion recognition and human-computer interaction due to the advantages of easy acquisition through wearable devices. PPG-based devices offer advantages of compactness, simple operation and high recognition rate. It compensates motion recognition using videos in continuous personal monitoring. In this paper, we have designed a system that paired the STM32F103C8T6 microcontroller with the MAX30101 PPG sensing platform and proposed a high-precision human motion recognition method based on PPG signals. The method extracted features using variational mode decomposition (VMD) and principal component analysis (PCA) for their advantages of adaptive mode decomposition and noise reduction, effectively enhancing signal quality for accurate classification. The features were then input into a CNN-Transformer network for the classification of three motion states, sitting, walking and jogging. The combination of CNN’s spatial feature extraction and Transformer’s long-range dependency modeling significantly contributed to the high accuracy in the recognition of motion states. Furthermore, the influences of light source wavelength and the PPG acquisition position were explored. It was shown that near infrared light sources and sensors being placed at the distal phalanges can provide the best signal quality and motion recognition results. The proposed method was compared to a variety of conventional methods and was shown to be superior. This research has demonstrated high precision human motion recognition based on PPG signals, and provided an idea of developing human motion recognition using PPG signals in wearable devices. This can be extremely useful in human motion science, exercise physiology, and health management, etc.</p>