Understanding deep neural networks performance for radar-based human motion recognition

Deep neural networks have recently emerged as a promising tool for radar-based human motion recognition. Their nonlinear structure makes them successful in classifying large-scale datasets. However, due to their complexity, it is difficult to interpret the classification results and identify pixels with the biggest impact on the classification score. In this paper, we investigate recently proposed linear-wise relevance propagation (LRP) method which finds relevant pixels within the image. Based on this method, it is possible to recognize pixels which contain evidence for or against the prediction made by a classifier. Experimental results demonstrate that the LRP method can be successfully applied to detect regions within the radar images responsible for distinguishing human motions.