Research On Deep Learning-based Human Behavior Recognition In Ultra-wideband Radar

Human behavior sensing and understanding are important parts of intelligent humancomputer interaction,which are promising in various fields such as emergency response,counterterrorism,and intelligent homecare.Diversified tasks require a high level on the accuracy,real-time and robustness.Fortunately,the electromagnetic wave provides a complementary input when the visible light cannot work.Therefore,the radar-based human activity sensing system can effectively tackle darkness,occlusion,and non-lineof sight conditions.In fact,the integration of statistical analysis and classical signal processing has been more and more prominent in the radar-based human behavior understanding field.This integration has effectively promoted the development of intelligent radar system and multi-mode biometric recognition.Specifically,deep learning,which is a branch of machine learning,broadens the content of pattern recognition.It is an end-to-end solution aiming to find the global optimal solution of given tasks.However,the conventional deep learning method is both data-hungry and resourcehungry.What’s worse,it assumes that the training and testing dataset accord to the same data distribution,which generally degrades the performance.To polish its performance in specific field,it is necessary to exploit the corresponding domain knowledge.In a nutshell,there are numerous challenges when we introduce deep learning to the ultra-wideband radar field.This paper focuses on the behavior classification and pose estimation of ultrawide radar human targets.The main contributions are as follows:To resolve the problem of limited training samples,we propose an L1 norm-based sparsity-driven transfer learning method and an unsupervised adversarial domain adaptation approach.Among them,the L1 norm-based sparsity-driven method evaluates the importance of each component of the network and prunes the whole network.This process accords to the difference between optical images and radar data.It effectively improves classification performance and reduces computation cost despite that the training set is highly limited.The adversarial domain adaptation method can quantitatively measure differences of various data sets.Moreover,it can classify human behaviors via unsupervised learning.This adversarial feature learning approach can improve the generalization ability of radar system for different signal-noise ratios and different clutter distribution environments.For the joint representation and recognition of micromotion signatures,we design a range-velocity-time point model to depict and recognize human activities.The designed point model is suitable for single channel ultra-wideband radar whose range resolution cell is smaller than the physical size of scattering centers.It jointly represents the rangevelocity-time information conveyed in the radar echo.It is shown that the proposed model is suitable for multi-target and noise-contaminated conditions.For the micromotion recognition,a hierarchical neural network is designed.Based on the network,an integrated process of in-distribution data classification and out-of-distribution data detection is proposed to realize human behavior recognition in an open set framework.For three-dimensional human pose estimation,we propose a novel method to estimate human pose using radial measurements.Owing to the lack of azimuth and altitude resolution,the single-channel radar measurement loses the structural skeleton of targets.To address this issue,we focus on one single motion,walking.Specifically,we detect this behavior by our previous works,explore the relationship between radial and three-dimensional target information,and estimate the corresponding three-dimensional pose with a dilated convolution network.Our research has been analyzed and evaluated on simulated human behavior backscattering echoes and ultra-wideband radar measurements.