Research On Human Behavior Recognition Of FMCW Radar Based On Machine Learning Method

Human behavior recognition technology has great application prospects in the fields of intelligent monitoring and biomedical research.This technology uses specific equipment or means to determine the state of human activity.Radar-based human body perception technology has become one of the current mainstream technologies for human behavior recognition due to its technical advantages such as good privacy protection,high range resolution,high measurement accuracy,and no fear of occlusion.This paper studies the human behavior perception technology based on Frequency Modulated Continuous Wave(FMCW)radar,and designs a perception system that realizes indoor human positioning and behavior recognition in two-person scenes.The specific research work is summarized as follows:(1)The realization scheme of indoor human body perception based on FMCW radar is studied.First,it analyzes the detection principle of FMCW radar’s distance,speed and angle,combined with the characteristics of indoor application environment,studies the elimination method of multipath interference,gives the use of background elimination algorithm to remove static interference,and adopts 2D constant false alarm An implementation scheme for filtering out dynamic multipath interference signals with Constant False-Alarm Rate(CFAR)detection algorithm.Finally,combined with Texas Instruments’ IWR1443 millimeter wave radar development board,an indoor human body perception system was built to verify the above-mentioned multipath interference cancellation algorithm,and analyze the influence of various parameter settings in CFAR on the accuracy of human body positioning.The actual measurement results show that This method can realize indoor human body positioning and effectively eliminate the influence of indoor multipath interference on the accuracy of human activity detection.(2)Research the processing method of radar point cloud data in indoor human body perception system.In order to realize the distinction between two-person point clouds,in view of the problem that the traditional Gaussian mixture model is sensitive to randomly selected initial clustering centers and the clustering results are unstable,an optimization algorithm based on the log-likelihood function is proposed to carry out the traditional model.Improved,the improved Gaussian mixture model improves the stability of clustering while ensuring real-time performance,and can better distinguish the active point clouds of two people.On this basis,in view of the sparse multiple reflection points in the radar frame,the data rate will affect the accuracy and real-time performance of the system.The method of multi-frame fusion to take the mean representative point is used to improve the signal-to-noise ratio while ensuring the algorithm Real-time.(3)The behavior recognition scheme of the indoor human body perception system is studied.Firstly,the method based on traditional machine learning is adopted to analyze the characteristics of falling,sitting and walking,and extracting discriminating features and inputting them into the random forest classifier.A good classification effect is obtained and the recognition accuracy is not high.Affected by the specific location of the target and the direction in which the action occurs.In addition,in view of the cumbersome process of manually extracting feature parameters in traditional machine learning,a behavior recognition method based on deep learning residual network models is further proposed.Through analysis and comparison,the optimal interpolation size,learning rate and optimizer of the data are optimized.Achieve a more ideal recognition effect.