TriBERT: Human-centric Audio-visual Representation Learning