Traffic vehicle detection plays important role in making decision about maintenance of a road section. However, the method to conduct it still used traditional approach, by means of surveyors being on the road and identifying vehicles for 40 hours, so it takes quite a long time and has the potential for human error to occur when identifying vehicles. In this research, a solution is formulated to identify vehicles using closed-circuit television (CCTV) and object detection methods based on deep learning. The dataset that used to train deep learning model were recorded in some of road section by our CCTV. Then, we annotate each object from given video frame based on defined classes. Then, all of the annotated frame divided in train and validation with percentage of 80% and 20% respectively. Train and validation dataset used for model training and test dataset used for evaluating best model weight and produce Average Precision, while best model weight also tested for show model performance and its Frame Per Second. We then compared the application of Faster-RCNN method with ConvNext v1 and Mobilenet v3 backbone in carrying out vehicle detection. Using 12 classes of vehicle in training and testing phase, test results based on evaluation dataset showed that ConvNext v1 backbone produced an average precision value of 0.81 while Mobilenet v3 backbone obtained a result of 0.3. As for the results of the Frame per Second (FPS) test, Mobilenet v3 backbone obtained an average FPS of 18 while Convnext v1 obtain 7. The results obtained indicated Faster RCNN backbone ConvNext v1 was an effective approach to obtain robust object detection while Faster R-CNN Mobilenet v3 backbone is effective for object detection in real time.