Abstract: Aiming at the complex scenes of traffic video surveillance and the characteristics of high real-time requirements,a vehicle detection algorithm based on“divide and conquer”thinking was proposed,which can run in real time under the CPU platform.By adopting guided filter on scene image and calculating the peak signal to noise ratio before and after filtering,the light conditions are detected.With multiple-model fusion method based on the classification result,the complex problem was separated into several sub-problems that can be solved with simple models.We then trained cascade Adaboost detection model for each of the sub-problems.With comparison to 4 deep-learning-based detection models on 15 groups of real surveillance test samples,this method shows high accuracy under complex scenes,and processes in real-time at 30 frames per second.