Classification of human target movements behind walls using multi-channel range-doppler images

Range-Doppler images represent one of the most informative radar data forms, providing range and frequency information. This study explores the performance of machine learning and deep learning techniques in classifying human activities behind walls using Range-Doppler images. Therefore, we input the HOG features of Range-Doppler images into various machine-learning approaches. Although the HOG feature enhances the performance of machine learning methods, we observe the superior performance of Convolutional Neural Network (CNN) architectures in a more complex scenario in which the number of target activity classes is higher. To obtain sufficient data for the CNN architecture, we combine the Uniform Linear Array (ULA) and stepped-frequency continuous-wave (SFCW) structures, enabling the acquisition of multi-channel data. The experimental results demonstrate both the improvement of the machine learning accuracy from 95.33% to 98.67% through the HOG + Range-Doppler approach and an approximately 6% enhancement in CNN performance achieved through the SFCW-ULA combination.