Through-Wall Human Motion Representation via Autoencoder-Self Organized Mapping Network

This paper addresses a problem of human motion representation for through wall radar exploiting an autoencoder-self organized mapping network. Motivated by the natural language processing, we interpret the echoes of the through-wall human motions into a simple integer sequence with motion “semantics” by regarding each range profile as a “word” and treating the multi-temporal-frame hidden human motion range profile data as a “sentence”, simplifying and maintaining the effectiveness of the through-wall human motion representations. Specifically, we firstly use an auto encoder network (AEN) with three dense layers to reduce the dimension and extract the features of each range profile. Then, a self organized mapping (SOM) network is employed to concatenate the trained AEN to establish the word-to-index mapping, transferring the temporal sequential range-profile data of through-wall human motions into an integer sequence which implies the “semantic” information of the motion. Finally, experimental data including four through-wall human motion types validate the proposed framework.