Recognition Of Human Motion Pattern Using Mmwave Radar

Human activity classification and recognition play an important role in the fields of safety,crime prevention,medical monitoring and so on.Limbs motion contains abundant micro-Doppler information in the body motion.Different human activities have different micro-Doppler distribution.These differences can be used to classify different human activities.Therefore,the deep learning method based on image domain has been widely applied in radar target classification and recognition.In most cases,the cost of acquiring radar images by measurement is very high,so difficult to construct large radar image database.And The key of train supervised deep learning is that large-scale labeled data is needed.To overcome these issues,the paper study the application of semi-supervised,unsupervised learning and data augmentation for human activity classification based on micro-Doppler signatures.Moreover,the performance of classification model is validated by the mmwave radar measurement data.In the experiment,the six kinds of human movements are selected for classification research,such as walking,running,boxing,jumping,standing and crawling.The main work of the paper as follows:(1)Dataset generation.The non-parametric human model is constructed by the movement data of each joint point of human body based on motion capture technology and the real situation of human motion restored.The radar echo of the human body can be obtained by the motion trajectory of each part of the human body.The joint timefrequency analysis method is used to generate the micro-Doppler spectrograms from the radar echo and the human activity training dataset is constructed.Similarly,mmwave radar is used to obtained the raw echo data of human activity.Combined with joint time-frequency analysis,the micro-Doppler spectrograms is generated to construct the test dataset,which is used to validated the performance of the classification model in practical scenarios.(2)Semi-supervised learning.In practice,most of the images are non-Euclidean structure data.Convolution Neural Network simplify the input image into regular Euclidean structure data for next operation.During feature extraction,for keep more effective information in micro-Doppler spectrograms,Convolution Neural Network is extended to the graph Convolution Neural Network to achieve semi-supervised learning,which improves the network classification performance and can reduce the scale of labeled data.(3)Unsupervised learning.Supervised deep learning has been successfully applied in human activity classification and pattern recognition based on radar images.Compared to natural images,the lots of manual labor and high costs involved in acquiring radar images.However,Generative Adversarial Networks(GANs)can generate large amounts of images similar distribution to the original images.The training dataset is augmented by GANs and unsupervised learning is introduced to study the classification of human micro-Doppler.