Detection and Recognition of Dynamic Human Postures in Non-Line-of-Sight Environment

Detection of dynamic human postures in non-line-of-sight (NLOS) region is implemented by using 77GHz linear frequency modulated continuous wave (LFMCW) radar. The recognition is performed by combining enhanced time-frequency posture features with transfer learning technique. By analyzing and denoising the radar echo signals from human targets in NLOS region, effective time-frequency information of human postures can be extracted. By applying the transfer learning technique with several pre-trained neural networks, all typical human postures can be successfully classified for the NLOS scenario. Experiment results show that the ResNet-50 network with transfer learning technique can provide the highest average classification accuracy of 97.92% for all the typical human postures.