Nowadays, the traffic situation is very complicated in Vietnam. Traffic jams happen frequently in densely populated or peak hours. So, it is necessary for an automatic warning system to police officers about traffic status in time and effectively. In this research, the system automatically estimates the vehicle motion rate and the number of vehicles. This system is useful for roads with lots of hard-to-distinguish traffic. The method is based on computer vision such as background subtraction and deep learning such as the CNN network model and the CenterNet object detection model. Experimental data are taken from videos in Vietnam with the view in front of theKien Giang and Da Nang hospitals. The achieved classification model results can predict with 91.7% accuracy on the test set, the precision of crowded road predictions is 81.9% precision and 70.4% recall. When the CenterNet model is applied to estimate the number of vehicles, the model reached 1.261 MAE. The speed of the system when it is run experimentally on hardware using Nvidia Geforce GTX 1070 GPU reached 4.6 FPS.