MESEN: Exploit Multimodal Data to Design Unimodal Human Activity Recognition with Few Labels