A vision-based occupant classification method utilizes a static classification routine for achieving highly accurate occupant classification, a dynamic detection routine for quickly responding to changes in occupant position, and a category state change routine for detecting potential changes in occupant category. The static classification routine determines an initial classification of the occupant, after which the dynamic detection routine monitors for changes in occupant position and the category state change routine monitors for a potential change in occupant category. If category state change routine detects a potential change in occupant category, the static classification routine is re-executed to re-classify the occupant. The category state change routine identifies a composite optical flow vector associated with occupant movement, and detects a potential change in occupant category when the composite flow vector crosses one or more pre-defined entry/exit motion gates.