Multi-human Separation Method Based on CNN-Mask-PIT

Abstract

Human activities classification based on micro-Doppler effect is a hot topic in the field of radar detection. However, when there are multiple humans, the motion mechanism of the multi-human target is complex, the components are mixed, and the micro motion features of each component are highly similar, which lead to difficulties in the classification of the multi-human activities. In order to overcome these problems, a multi-human separation method based on convolution neural network Mask generation and permutation invariant training (CNN-Mask-PIT) is proposed in this paper. In this method, the multi-human target is separated on the time-frequency (TF) images. The Mask matrixes are generated by CNN, and then the generated multiple Mask matrices are multiplied with the original TF image to obtain the separated TF images, so as to realize the effective separation of multi-human separation. The permutation invariant training (PIT) is introduced to solve the label arrangement problem by minimizing the separation error. Finally, the effectiveness and robustness of the proposed method are verified on the multi-human TF images generated based on motion capture data.