Font Size: a A A

Research On Human Motion Recognition Method Of UWB Radar Based On Deep Learning

Posted on:2024-08-21Degree:MasterType:Thesis
Country:ChinaCandidate:G WangFull Text:PDF
GTID:2568307079966329Subject:Electronic information
Abstract/Summary:PDF Full Text Request
The ultra-wide band stepped-frequency radar can acquire the target high-resolution range map and multi-frequency time-frequency spectrogram simultaneously.Using range map and multi-frequency spectrogram fusion for robust human motion recognition to sense the motion state of the target,it can be applied in the fields of anti-terrorism combat,law enforcement apprehension,health monitoring and human-computer interaction.In this thesis,based on the ultra-wide band stepped-frequency radar platform,the research is carried out for human motion recognition with fixed length input and con-tinuous input respectively.For the human motion recognition with fixed-length input,3-dimensional electromagnetic echo simulation,multi-frequency temporal spectrogram human motion recognition based on frequency attention mechanism,and multi-domain human motion recognition method with multi-frequency spectrogram and range map fu-sion are carried out.For continuous input radar human motion recognition,research on continuous human motion recognition method based on attention encoder-decoder and continuous human motion recognition method based on convolutional recurrent network(CRNN)and connectionist temporal classification(CTC)loss function are carried out.The main work of this thesis is as follows:1.The human motion baseband echo signal model of ultra-wide band stepped-frequency radar is derived.And for the problem that the traditional numerical simulation method of radar human motion does not consider the attenuation of electromagnetic wave energy with distance,an ultra-wide band stepped-frequency radar human motion baseband echo electromagnetic simulation method based on motion capture data and 3-dimensional(3D)finite difference time domain(FDTD)method is proposed,which solves the problem that the human motion echo of numerical simulation cannot reflect scattering intensity infor-mation of human target at different distances.2.The human motion recognition method based on the frequency attention mech-anism is proposed for the effective utilization of the time-frequency spectrogram infor-mation of multiple frequencies of ultra-wide band stepped-frequency radar.It uses the attention mechanism to fully exploit the human motion frequency attributes of multi-frequency spectrograms and fully utilize the frequency complementary features between the time-frequency spectrograms of multiple frequencies in order to improve the accuracy of human motion recognition.The experimental results show that the recognition method using multi-frequency attention mechanism achieves 96.82% accuracy for 9 types of hu-man motions,which is better than the 95.38% accuracy of single frequency spectrogram method and 96.27% accuracy of direct fusion of multi-frequency spectrograms.3.In response to the problem that the human motion recognition method using only multi-frequency spectrograms has similar features in the temporal frequency spectrograms of some categories of human motion,and the recognition accuracy is significantly lower than that of other categories,a multi-domain human motion recognition method using multi-frequency time-frequency spectrograms and range map fusion is proposed.It makes full use of the complementary characteristics of multi-domain of time-frequency and range domains to further improve the accuracy of human motion recognition.The recognition accuracy of human motion with 9 types of radar reaches 98.79%.4.For the continuous input radar human motion recognition problem,continuous human motion recognition datasets with different sequence lengths are produced.The time-frequency spectrograms of continuous input are modeled as image sequences,and a continuous input human motion recognition method based on encoder-decoder and atten-tion mechanism is proposed.The attention mechanism can make full use of the contextual information of convolutional features to improve the accuracy of continuous human mo-tion recognition,and finally the accuracy of continuous recognition of 9 types of human motion reaches 98.55%5.The attention weight calculation for the encoder-decoder model requires a com-plete sequence,leading to it cannot be computed online in real time and the training se-quence of the encoder-decoder model needs to be strictly aligned with the labels according to the convolutional receptive field,which brings huge workload to the training data label-ing.In this thesis,a continuous human motion recognition method based on CRNN and CTC loss function is proposed,which training process does not require alignment between spectrogram input and label,and also can perform streaming real-time online recognition.
Keywords/Search Tags:Ultra-wide Band Radar, Human Motion Recognition, Multi-frequency, Attention Mechanism, Continuous Recognition
PDF Full Text Request
Related items