When detecting the human targets behind walls, false detections occur for many systematic and environmental reasons. Identifying and eliminating these false detections is of great importance for many applications. This study investigates the potential of machine learning (ML) algorithms to distinguish between the human and human-like targets behind walls. For this purpose, a stepped-frequency continuous-wave (SFCW) radar has been set up. Experiments have been carried out with real human targets and moving plates imitating a regular breath of a healthy human. Unlike conventional methods, human and human-like returns are classified using range-Doppler images containing range and Doppler information. Then, the histogram of oriented gradients (HOG) features of the range-Doppler images are extracted, and the number of these features is reduced by principal component analysis (PCA). Finally, popular ML algorithms are executed to distinguish the human and human-like returns. The performances of the ML algorithms are compared for both range-time and range-Doppler images with or without HOG features. Experiments have indicated that the HOG features of the range-Doppler profiles provide the best results with the support vector machine (SVM) classifier with an accuracy of 93.57%.