Research On Dangerous Driving Behavior Recognition Method Based On FM Continuous Wave Radar

With the popularity of vehicles,safe driving of automobiles has received widespread attention.In response to the prevalence of dangerous driving behavior during the driving process of car drivers,real-time warning of dangerous driving behavior is required in a moving closed space.A huge amount of human behavior sensing data will be collected during vehicle driving,and due to the special requirements of driving scenarios,a precise,fast,and anti-interference solution is needed.FM continuous wave radar is widely used in various scenarios of perceptual granularity with the advantages of high recognition accuracy and not be easily affected by the environment.In this paper,based on the frequency modulated continuous wave(FMCW)signal generated by millimeter wave,we extract multi-dimensional Doppler features,design different data processing models and behavior recognition classification methods,and realize driver behavior sensing with high accuracy,high robustness,and low latency.This thesis focuses on an in-depth study of dangerous driving behavior recognition in multipath scenarios,and the main work and contributions are as follows:(1)For the problem of serious multipath effect caused by small space and dense personnel in the vehicle,an MDP(Micro Doppler multi-Path)multi-order multipath superposition signal processing model based on micro-Doppler spectrum information is proposed.In addition to the target driver movement information in the car there are also unrelated personnel and the reflected signal of the metal components in the car,resulting in unrelated noise interference,noise interference will cause the echo phase shift,radar on the signal different distances generated by different orders of echo signal analysis and processing.Firstly,matched filtering is used to maximize the elimination of positive and negative mismatches to cancel the interference and solve the distance Doppler coupling phenomenon in signal processing.Then,the multipath information is modeled,and the different distance echoes are divided into zero-order,first-order and multi-order reflections,and the first-order reflections are modeled and calculated by different reflection methods to achieve the effect of non-target reflected clutter through the model to enhance behavioral characteristics.The MDP model effectively reduces the negative impact of multipath effects on behavioral recognition in the complex environment of the vehicle.(2)For the problem of perceiving fine-grained dynamic dangerous driving behaviors,a micro-Doppler lightweight two-channel lateral fusion processing network is proposed.First,the dangerous behavior Doppler feature action information is obtained by baseband processing and fast Fourier transform,and the optimal feature combination is selected for upper limb feature recognition,and the clutter is filtered using a multilayer filter.Secondly,two parallel network paths with different temporal velocities are driven by simultaneous sequential inputs,with the low-frequency sampling channel capturing tiny slow target signals and the fast sampling network channel capturing target signals of larger waveforms.Finally,2D-B-Alex Net is used to extract inter-frame coherence information of target actions,and the fast and slow paths are classified based on the horizontal dual-channel fusion of time-series lateral information,which effectively improves the convergence of behavior recognition and the overall accuracy of data set recognition.(3)To address the problems of Doppler feature data redundancy and dynamic articulation of multi-feature fusion,an improved attention mechanism low-rank multimodal LMF-GRU model is proposed to improve the reliability of behavior recognition.Firstly,the Range-Dopple Matrix(RDM)method is used to map the parsed data in different dimensions in consecutive frames with fixed time windows to generate three feature vectors,which are the key micro-Doppler feature vector,the hidden radial velocity feature vector and the hidden amplitude feature vector.Secondly,the three features are tensor fused by a low-rank multimodal approach,which makes the three low-dimensional data normalized to a high-dimensional tensor data representation,which reduces the complexity while the high-dimensional tensor can completely represent the time window of dangerous driving behavior.The LMF-AR-GRU model solves the problem of normalizing complex Doppler information and effectively improves the overall performance of the system.