Lane departure warning system(LDWS) is one of the main functions for the vehicle active safety system. This paper researches some critical techniques of LDWS. The road model is made as the core and detect lanes based on the model driven approach. In order to detect curve, the region of detection is divided into multiple sub regions, then lane marks are detected in each region independently. For speeding up the detection and decrease error, the size of each sub region is changed based on the detection result and fitting values of the road model in each frame. For increasing the detection stability, the Sobel filter is adopted for nearby sub regions and Gabor filter is adopted for distant sub regions. Furthermore, the Gabor filter is set adaptively based on the road geometry. Experimental results shows the system can accurate detect the curve and predict lane departure range precisely. Furthermore, the system has good robustness in the noisy environments.