Research On Pedestrian Target Detection And Identification Technology Based On Radar Microdoppler Signatures

The demand for pedestrian detection and identification is growing by the day,thanks to the advancement of intelligent driving.Radar sensors are less affected by environmental conditions(such as fog,heavy rain,thick smoke,etc.)than other systems such as cameras and Lidars,and can guarantee high performance and all-weather work at a relatively low cost.As a result,radar sensors are increasingly being used in automotive pedestrian identification systems,attracting the interest of academics.When the radar target is a pedestrian,its arms and legs swinging periodically throughout time generate distinct micro-Doppler signatures,which can be utilized to distinguish between various walkers.Time-Doppler spectrogram and signal statistical features are two types of micro-Doppler signatures extracted in this thesis,which contain the micro-motion information of pedestrians walking freely in two rooms.In this thesis,two methods for pedestrian identification based on radar micro-Doppler signatures are proposed,and the effectiveness of these two methods in pedestrian identification in a dataset closer to reality is experimentally demonstrated.The first is a multi-characteristic learning(MCL)model with clusters,which is also proposed in the thesis to jointly learn two categories of pedestrian micro-Doppler signatures(i.e.,time-Doppler spectrogram and signal statistical features)and fuse the knowledge learned from each cluster into the final decision.The experimental results show that compared with other studies,the MCL model achieves higher accuracy and is more stable for pedestrian identification,which makes the MCL model more practical.The second method is a new convolutional neural network-based method,called Multi-Scale CNN(MS-CNN),to obtain features at multiple scales of the time-Doppler spectrogram.It extracts shallow features in low-level multi-scale blocks by using convolution kernels of multiple scales,then extracts deep features in high-level multibranch blocks and fuses multi-branch embedding features.The experimental results show that the MS-CNN method has a significant improvement in accuracy compared with other commonly used pedestrian identification methods.This thesis leverages a public pedestrian dataset,IDRad,to propose a solution to the real-world pedestrian detection problem.In contrast to the datasets used in other studies,in this dataset,pedestrians can freely move in both rooms without being restricted in their walking direction,which makes the two methods proposed in this thesis more suitable for real environments.The accuracy of these two methods on the test set can reach 87.63 and88.57,respectively.