It is principal to detect lane robustly and rapidly for intelligent vehicle based on the information of road marking or road region. The road image is divided into tow parts called near area and far area based on pre-knowledge and human visual experience. A linear model is adopted to fit the lane marking in near area, while in far area, the lane marking with the lane model is switched between linear model and cubic curve model. Combined with the gradient value, gradient direction and gray information, discriminant function of the probability is derived. Then the improved ParticleSwarm Optimization (PSO) algorithm combined with genetic algorithm operators is used to quickly search the optimal model parameter of the discriminant function to implementation lane detection. The results of the real road image experiment show the proposed method can robustly and rapidly detect the lane markings even if there are some interference factors in the road such as shadow, non-uniform illuminance, vehicle barrier and soiled lane boundaries.