In order to process multitasking driver behavior effectively, an improved driver cognitive behavior modeling method of ACT-R is proposed in this paper. The manual and visual modules of ACT-R are concatenated directly to cope with human subconscious/unconscious behavior. A parallel processing method is also proposed to mimic the parallel reaction style of a given cerebral area of human brain's reaction to the physical characteristics of the stimulus. Driver behavior assorting and risk level ranking methods are applied to improve the model's executive efficiency. The simulation results demonstrate that the ACT-R cognitive architecture is improved significantly with theses proposed improvements.