This　letter　presents　a　novel　communication-efficient　and　decentralized　approach　for　data　analytics　in　connected　vehicles.　We　extend　the　paradigm　of　federated　learning　(FL)　to　enable　decentralized　on-vehicle　model　training　without　a　central　server.　To　improve　communication　efficiency,　we　design　a　federated　regularized　nonlinear　acceleration-based　local　training　scheme　to　reduce　the　communication　rounds　and　a　random　broadcast　gossip-based　mechanism　to　decrease　the　complexity　per　iteration.　Experimental　results　demonstrate　that　our　approach　significantly　reduces　the　communication　cost　compared　to　general　gradient　descent　and　momentum-based　FL　solutions　and　is　promising　for　efficient　data　analytics　in　autonomous　vehicle　environments.