A Deep Learning Method For Vehicle Radar Pedestrian Recognition Based On Micro-Doppler Signatures

Pedestrian recognition is one of the most important issues in the field of autonomous driving.Compared with sensors such as cameras and ultrasonic radar,the on-board millimeterwave radar is not affected by harsh weather such as haze,rain,and snow,as well as the intensity of ambient light.It can measure the distance,speed,and motion direction of targets,and has high accuracy in target recognition.Therefore,millimeter-wave radar is widely used as a sensor for pedestrian recognition.The micro-Doppler effect provides theoretical support for pedestrian recognition using millimeter-wave radar.The swinging of hands and legs relative to the torso during walking produces the Doppler effect,and each person has unique physical characteristics and walking habits,so the echo signal contains unique micro-Doppler signatures.This thesis extracts two types of micro-Doppler signatures,time-Doppler maps and physical features,from the radar signal,and proposes a deep learning method based on microDoppler signatures for pedestrian recognition and a data enhancement method based on generative adversarial networks.The effectiveness of these two methods in pedestrian recognition is demonstrated through experiments.A deep learning network model based on multi-task learning is proposed for the pedestrian recognition problem based on micro-Doppler signatures.The model jointly learns time-Doppler maps and physical features extracted from the map,and integrates the knowledge learned from the two types of micro-Doppler signatures as the final recognition result.Experimental results show that the model proposed in this thesis has higher recognition accuracy and stable recognition results compared with other deep learning models for pedestrian recognition.For cases where the data samples are insufficient,a data enhancement method based on generative adversarial networks is proposed.Since there is temporal correlation between consecutive range-Doppler maps,a generative adversarial network RDGAN(RangeDoppler Generative Adversarial Networks)that can learn both spatial and temporal correlations is designed to generate range-Doppler maps that are similar to real data,thereby expanding the dataset and improving the model’s generalization ability.Experimental results show that the multi-task learning model with data enhancement can achieve higher recognition accuracy.This thesis uses a dataset that is closer to real scenarios,which includes radar data collected from people walking freely in two different rooms at different times.The recognition accuracy of the multi-task learning model on the test set is 88.86%,and it can reach91.52% after data augmentation with generative adversarial networks,the results indicate that the method proposed in this thesis provides an effective and more realistic solution to pedestrian recognition problems.