Research On Radar Based Human Activity Recognition In Complex Scenes

Radar based human activity recognition in complex scenes can be applied to identify and predict the behavior states and intentions of sheltering terrorists in buildings,to perceive the real-time situation of the battlefield and the predicted dangers,to keep privacy to know the recovery of non-contact monitoring patients,and to detect unexpected situations such as falls of the elderly,which provides strong technical support for rescue programs and operational decisions.However,due to wall shielding,clutter and interference in space,multiple targets within the radar detection range and possible overlap in time and characteristics,the continuous behavior of a series of behaviors is diverse and time-varying,and relatively weak in-place behavior requires a high resolution of the radar in complex environments or tasks,making human behavior recognition a challenging problem.In this paper,around the above problem of real-time recognition of continuous inplace behavior of multiple people with wall covering in complex scenes,the research on Modeling of continuous echo,acquisition and processing of in-place behavior data,recognition of in-place behavior of multiple people,real-time recognition of continuous in-place behavior of multiple people are carried out,and the experimental verification is carried out by simulation and real-time measurement data.The specific work is as follows:1.A signal processing method based on signal transformation and fusion is proposed for radar wall covering human in-place behavior.Various data forms are obtained by signal transformation methods such as time-frequency analysis,etc.A clutter suppression method is designed to suppress all kinds of clutter that may exist in it,and a variety of fusion methods are used to enhance information about in-place behavior and form a more comprehensive representation,solved the difficult data representation problems caused by complex wall shielding electromagnetic environment and relatively weak in place activity,realized the complex scenarios of radar effectively describe human activities.2.A data enhancement method based on the combination of single sample generation and fast filtering is proposed.A large amount of data is generated from a single sample through multiscale generation antagonistic network,and then the effective generated samples are filtered by a migration learning based classifier,which initially solves the problem of small samples caused by difficulties in radar data collection.3.A single-stage multi-person behavior recognition method with the integration of split recognition and a two-stage multi-person behavior recognition method with the combination of split recognition and recognition are presented.One-stage method constructs a cascade recognition architecture based on split attention and mask to segment and identify multiple targets at the same time.Two-stage method first extracts singleperson data through density clustering or edge detection,and then extracts multi-domain features through ordered loop neural network to get recognition results,effectively solving the problem of stable and accurate in-place behavior recognition in multi-person scenes.4.A lightweight network based on random circulation of space-time characteristics is proposed to realize real-time recognition of human continuous in-place behavior.Improvements based on sub-manifold sparse convolution,single-headed self-attention and phase fusion are made for possible instability factors,which theoretically enhance the accuracy and robustness of human continuous in-place behavior recognition.