Real-Time Moving Vehicle Counting and Speed Estimation Toward Efficient Traffic Surveillance

This paper presents a Spatial-Temporal Diagram (STD) algorithm for real-time vehicle counting and speed estimation in camera-based traffic surveillance. The algorithm consists of four main steps: STD graph generation highlighting vehicles as peaks, graph refinement using Gaussian Mixture Model likelihood optimization, peak detection through RANdom SAmple Consensus model fitting, and traffic parameter computation. Testing on over 11 million video frames from diverse sources, including 511 highway cameras, NVIDIA AI City Challenge, and Next Generation Simulation datasets, demonstrated the algorithm’s robustness across varying illumination, weather conditions, and road infrastructures. The algorithm achieved average accuracies of 95.4%, 96.9%, and 96.1% for Precision, Recall, and F1-Score, respectively, outperforming traditional deep learning methods while requiring less computational resources.