Indoor Human Behavior Recognition Method Based on Wavelet Scattering Network and Conditional Random Field Model

Ultra-wideband (UWB) through-the-wall radar (TWR) can be utilized for indoor human behavior recognition via micro-Doppler information. However, faced with weak micro-Doppler features and low accuracy, conventional recognition method does not perform well in shielded environment. To address these problems, this paper proposes an indoor human behavior recognition method based on wavelet scattering network and conditional random field model (TWR-WSN-CRF). In the proposed method, wavelet scattering network, speckle reducing anisotropic diffution (SRAD) with weighted guided image filter (WGIF) are employed for feature enhancement and noise suppression, and the signal-to-noise ratio (SNR) is improved. Then, human behavior recognition network based on conditional random field model is developed to extract global and local features from wall, target, noise subspace obtained by singular value decomposition. Finally, multi-layer perceptron (MLP) model and weighted majority voting (WMVE) method are used for fusion decision. The effectiveness of the proposed method is verified by experiment. The results show that compared with other methods, the proposed human behavior recognition method achieves highest recognition accuracy with 96.25% on the validation data set.