Novel Machine Learning Driven Framework for Dynamic Perspective Correction in Vision-Based Traffic Speed Estimation

Accurate speed estimation is crucial for legitimate enforcement and effective traffic management. However, in computer vision-based systems, perspective distortion often leads to significant estimation errors, particularly under heterogeneous traffic conditions prevailing in developing countries. Traditional methods, such as the uniform scale factor (USF) approach, fail to correct for these distortions adequately, frequently resulting in unjust penalties when used in enforcement cameras. The present study introduces a novel machine learning-driven dynamic scale factor (DSF) framework for correcting perspective errors in computer vision-based speed estimation. The proposed method computes the DSF by incorporating intrinsic camera parameters and geometric parameters based on camera orientation. An annotated database adaptable to heterogeneous traffic conditions in line with the vehicle categories specified in the Highway Capacity Manual has been developed. A dedicated machine learning algorithm is developed to estimate speeds using the DSF framework. The methodology is validated on a test stretch in Kerala State, India. Vehicle speeds were estimated using the proposed DSF approach and compared with estimates obtained using USF, segmented scale factor (SSF), and ground-truth data. The results demonstrate that the DSF method yields speed estimates that are significantly closer to the ground truth, compared to USF and SSF methods. Crucially, the DSF approach eliminates the need for multiple field-based reference lengths and does not require individual camera calibration, making it highly scalable. Additionally, the method ensures reliable speed data for advanced traffic management applications. The proposed framework is adaptable and enhances the accuracy and fairness of speed estimation across diverse operational scenarios.