In recent years, convolutional neural networks (CNNs) have been increasingly used for classifying radar micro-Doppler signatures of various targets. However, obtaining large amounts of data for efficient CNN training in defence and surveillance scenarios can be challenging. Therefore, designing techniques that maximize the use of available samples is critical. In this paper, we propose an approach built on the hypothesis that certain classes of radar spectrograms, such as those used for discerning armed from unarmed walking personnel, do not have information about the class encoded in the trajectory. Therefore, our method entails segmenting each input spectrogram into individual frames that correspond to a distinct step of human locomotion. Subsequently, we classify each segment independently and combine the resulting classification scores to obtain the final score for the entire spectrogram. As a result of this segmentation, the size of the training set is increased, whereas the dimensions of each sample - and therefore the number of parameters in the classifier - is decreased, reducing the risk of overfitting. Our experimental results demonstrate the effectiveness of our approach and its potential to enhance CNN-based classification of micro-Doppler signatures.