A Deep Multi-task Network for Activity Classification and Person Identification with Micro-Doppler Signatures

In prior work, both person identification and activity classification in radar have been investigated. In this paper, we go one step beyond separate radar-based person identification and activity classification by proposing a deep multi-task network to complete the two tasks simultaneously. MOCAP dataset, from Carnegie Mellon University, is used for micro-Doppler signatures simulation. Six activities performed by six persons are adopted. 99.54% activity classification accuracy and 96.25% person identification accuracy is achieved with the network. To understand the networks impact on performance, ablation study is performed and experiment shows that each component in the network can facilitate the multi-task learning.