Abstract: Aiming at the low accuracy of video traffic statistics and vehicle type recognition, A novel multi-vehicle traffic flow detection method based on target tracking and convolutional neural network (CNN) transfer learning is proposed. Firstly, the CNN pre-training model MobileNet is used to learn the vehicle sample migration of the experimental scene, and the vehicle classification model is obtained. Then extracting the moving vehicle from each frame of the video, analyzing the center point of the adjacent frame vehicle, and establishing a target tracking model; Finally, the newly detected vehicles are input into the vehicle classification model and counted according to the three types of cars, trucks and passenger cars. The experimental results show that the method is compared with the detection methods based on virtual coils and support vector machines. The accuracy of traffic flow detection was 98.7%, an increase of 3%; The average accuracy of the vehicle classification was 96.8%, an increase of 15%.