Semantic annotation for complex video street views based on 2D-3D multi-feature fusion and aggregated boosting decision forests