A new deep learning architecture proposal for vehicle speed detection from surveillance cameras

This research presents a novel deep learning-based method for accurate and efficient vehicle speed prediction in intelligent transportation systems. Unlike traditional approaches focused on object detection and tracking, the proposed method employs a unified method that directly learns vehicle speed by integrating spatio-temporal information from video data. Vehicle motion is represented as trajectories in the spatio-temporal domain, where the slope of these trajectories directly determines the vehicle's speed. The framework's design enhances prediction accuracy and model robustness by integrating spatial and temporal cues seamlessly. In this context, the model has been developed as an extension of the popular YOLOv8 object detection framework and optimized for vehicle speed prediction in various traffic scenarios. The model is trained on extensive surveillance video data, effectively learning the complex relationships between visual features and vehicle speeds. Experimental results demonstrate the superior performance of our method, achieving a mean average precision of 99%, a low mean absolute error of 1.34 km/h, a mean absolute percentage error of 2.50%, and a root mean square error of 1.80 km/h. Additionally, proposed model achieves an inference speed of 150 frames per second, making it a suitable solution for real-time traffic analysis applications. This study highlights the effectiveness of a deep learning approach that directly learns speed from spatiotemporal data, offering an innovative solution for vehicle speed prediction. The proposed model provides a new approach by optimizing both accuracy and processing speed, contributing to advancements in traffic monitoring and congestion management.