Vehicle Speed Estimation Using Consecutive Frame Approaches and Deep Image Homography for Image Rectification on Monocular Videos

In intelligent transportation systems, vehicle speed estimation plays a vital role in traffic monitoring, speed enforcement, and autonomous vehicles. Therefore, the authors proposed a vehicle speed estimation method composed of pipelines: homography transformation using a deep image homography transformation network, vehicle detection by YOLOv8, tracking by ByteTrack, speed estimation in consecutive frames, and regression using Multiple Linear Regression. Homography transformation was utilized to rectify the monocular video view image into a bird’s-eye view with a constant pixel-per-meter value. In this scheme, the authors proposed a method for estimating the speed in consecutive frames using statistical and machine learning approaches by comparing several experimental schemes to determine which method is better for estimating vehicle speed, such as vehicle speed estimation results in Root Mean Square Error (RMSE) and Mean Absolute Error (MAE), traffic conditions, vehicle directions, and state-of-the-art studies. A monocular video DeaSpeedDataset was generated to evaluate the proposed system method by comparing the predicted speed value with the ground truth value obtained by the speed gun. This dataset contains 2408 vehicles divided into 20 conditions, four setups, and two locations. Furthermore, the best RMSE was approximately 2.37897 km/h, and the MAE was approximately 1.68977 km/h.