Enhanced Vehicle Speed Estimation Using Regression Models and Frame-by Frame Analysis with Deep Homography Transformation

Vehicle speed estimation is crucial in intelligent transportation systems.Applying a vision-based sensor to estimate vehicle speed without additional sensors is more efficient and cost-effective.Therefore, this study proposed vision-based speed estimation methods: image rectification using deep homography, vehicle detection and tracking using You Look Only Once Version 8 (YOLOv8) and ByteTrack, and speed estimation using a regression model, evaluated in DeaSpeedDataset.We improved the estimated speed results in the regression part and compared six regression models: Decision Tree, Multi-Layer Perceptron, Support Vector Regression, Ridge, Lasso, and ElasticNet, with the prior study using Multiple Linear Regression.The proposed study improved compared with the previous research of Root Mean Square Error (RMSE) by 2.87% for average results and 4.43% for best results by ElasticNet, with RMSE of 2.27.For Mean Absolute Error (MAE), the improvement was 0.15% for average results and 14.60% for best results by Lasso, with MAE of 1.44.