Physics-constrained multi-parameter lane-level traffic prediction from surveillance videos

Short-term traffic flow prediction is crucial for enhancing the control efficiency of intelligent transportation systems. However, most existing methods are limited to single-parameter prediction at the road-segment level, failing to exploit the complex interactions among multiple traffic parameters and their spatial distribution across lanes, which constrains both prediction accuracy and potential for fine-grained management. To address this, a video-driven lane-level multi-parameter short-term prediction framework that integrates computer vision and optimization theory is proposed. The framework follows three coherent steps: first, high-precision simultaneous extraction of lane-level traffic flow, speed, and density from surveillance videos; second, a bidirectional long short-term memory network that jointly models temporal dependencies and inter-lane spatial relationships to produce short-term forecasts; third, a measurement-adjustment-theoretic lane-level joint optimization model that collaboratively corrects the initial predictions to ensure physical consistency (e.g. flow conservation). Experimental results demonstrate that the framework effectively improves prediction performance. On multiple real-world road sections, the mean absolute percentage errors for flow, speed, and density reach 2.89%, 1.01%, and 2.72%, respectively, outperforming state-of-the-art baselines significantly. This study provides a systematic technical solution for fine-grained, reliable lane-level traffic state perception and prediction, laying a solid foundation for advanced applications such as dynamic lane management and real-time congestion mitigation.