Vision-centric Semantic Occupancy Prediction for Autonomous Driving