Aiming at the lane change behavior recognition requirements for vehicle active safety system, natural driving test in real road were carried out and different parameters related to lane change behavior were collected synchronously. Firstly, parameters were processed with Kalman filter to increasing the potential relevance among sample data. Then, SVM model was established for lane change recognition. Lastly, data normalization, principal component analysis method, and bayesian network were adopted to optimize the SVM model. The recognize rate of lane change with 1.2 second time window increased from 93.9% to 98.7% by using these optimization measures. It can be meet the requirements of effectiveness and real-time for vehicle active safety system, such as lane change warning system or lane departure warning system.