Traffic congestion is a daily occurrence in most urban of Vietnam, because of many reasons: an increasing number of vehicles, the rush hours (time to go to school, to work, after school, after work). To partly avoid these situations, it is necessary for a support system to provides information of traffic flow on key roads with traffic cameras at junctions, crossroads, in front of hospital gates, markets ... Estimating the traffic density from the traffic cameras helps the functional forces divide and divert traffic in time. In this research, the Faster R-CNN and YOLOv4 models is used to detect vehicles in traffic on the road, thereby estimating the density and sending alerts to the user. Experimental data were taken in front of Kien Giang hospital’s gates, including 1260 images for training, 360 images for validation, and 3 sets of images for testing. The first set of 180 images has the same camera angle as the training set. The second set and the third set each contains 100 images with different camera angles and different camera angles from the training set. The results are about 92% of accuracy in detection of vehicles for the first set (2 models); 75% and 35% for the second set; 85% and above 20% for the third set and above 20% for the third set.