Multi-Classification Algorithm for Human Motion Recognition Based on IR-UWB Radar

• Published in: IEEE sensors journal Vol. 20; no. 21; pp. 12848 - 12858
• Main Authors: Qi, Rui, Li, Xiuping, Zhang, Yi, Li, Yubing
• Format: Journal Article
• Language: English
• Published: New York IEEE 01.11.2020; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Accuracy; Algorithms; Artificial neural networks; Classification; Classification algorithms; convolutional neural networks (CNNs); Doppler radar; Doppler shift; Doppler shifts feature; Feature extraction; Human motion; Human motion recognition; IR-UWB radar; k-NearestNeighbor (KNN); Legged locomotion; Motion perception; power spectrum feature; Recognition; Sensors; Signal processing; Spectrograms; Ultrawideband radar
• ISSN: 1530-437X; 1558-1748
• DOI: 10.1109/JSEN.2020.3000498

In this paper, a multi-classification algorithm for human motion recognition based on Impulse-Radio Ultra-wideband (IR-UWB) radar is presented. The algorithm includes three parts. First, the k-NearestNeighbor (KNN) algorithm is used to classify the radial features of pre-processed signal to determine the subject's radial displacement direction. Then, the power spectrum feature extraction algorithm and Doppler shifts feature extraction algorithm are proposed to extract and visualize the characteristics from the different categories classified by the first part. Finally, the feature spectrograms obtained by the second part are sent into Convolutional Neural Networks (CNNs) for training and testing to realize the recognition of human motions. To verify the performance of proposed algorithm, dataset was created from 15 persons including 12 kinds of motions. The Five-Fold Cross Validation was conducted to calculate the recognition accuracy. As a result, the average accuracy of judging the radial displacement directions of subjects was up to 99%. Furthermore, the average accuracy of estimating the motions of subjects reached 98%. Experiments have proved that the proposed algorithm can achieve high recognition accuracy in daily human motions and is feasible in a variety of test environments.