Micro-Doppler Image Denoising And Human Activity Recognition Based On Deep Learning

Micro-Doppler radar based human activity recognition has been widely applied in many fields such as disaster rescue,smart home,automatic driving and security monitoring,which has great research value.Aiming at the practical application,this thesis studies the micro-Doppler image denoising and robust human activity recognition.Then,denoising model and robust human activity recognition model are porposed respectively.The performance of proposed models are verified by simulated and measured radar data.The main research contents and innovative work are described as follows.In order to reduce the noise in micro-Doppler image,this thesis proposes a denoising model based on Generative Adversarial Networks(GAN).To avoid the information loss in the neural network feed-forwarding,an activity feature representing network is proposed,which can introduce additional activity category information into the model.A novel loss function,which is constructed from three aspects: pixel distribution,semantic discrimination and human activity characteristics,is proposed to effectively improves the model performance.Denoising performance of the proposed model is verified on both simulated and measured radar data.The results demonstrate that the model has good denoising performance and can adaptively reduce the noise of various intensities.Based on the Range-Doppler(RD)sequence,a human activity recognition model composed of Convolutional Neural Networks(CNN)and Recurrent Neural Networks(RNN)is proposed.The model takes the Range-Doppler(RD)sequence as input and employs an improved CNN to extract RD features.Then,the RD feature sequence is taken as the input into the RNN,and bidirectional semantic is learned to achieve the human activity recognition.In order to further improve the recognition accuracy,this thesis proposes a novel attention mechanism,which assigns attention weights to the RD feature sequence based on the time series characteristics of the data.The proposed recognition model is validated by the measured radar data.The results demonstrate that the model not only has better recognition accuracy,but also has higher robustness to micro-Doppler radar signal of variable duration.In the end,the proposed denoising model and activity recognition model are integrated into a denoising-recognition system,and the activity recognition performance of the denoising-recognition system is verified on noisy RD sequences.The experimental results show that the system is not only robust to noise but also can accurately recognize the signal of variable durations.