There are many ways to detect vehicles’ speed these days, which can be categorized into two approaches: a non-computer-vision-based and a computer-vision-based. In this paper, we propose a computer-vision-based approach using YOLOv4 and XGBoost Regression. To predict vehicles’ speed efficiently, we use YOLOv4 for vehicle detection and XGBoost regression for speed prediction. In order to get the best speed prediction, we build our dataset by recording the local traffic and measuring their speed using a speed gun. From those traffic videos, we detect vehicles by using YOLOv4 to generate its bounding boxes. From the bounding boxes, we can extract its coordinates relative to the screen, the distance and the angle between the two centroids, and the time it takes from point A to point B. This information will be our features, and the speed from the speed gun will serve as the target to train our XGBoost regression model. In this paper, we conduct several experiments using various features to get the best model. Our experiments conclude that our speed prediction approach using YOLOv4 and XGBoost regression has a very high performance regarding to the ground truth with an MAE of just 2 km/h.