ChattSpeed: Toward a New Dataset for Single Camera Visual Speed Estimation for Urban Testbeds

Accurate speed estimation of vehicles is crucial for applications like congestion management and speed limit enforcement to reduce accidents. Despite its relevance, datasets for visual speed estimation are limited. In this study, we propose a deep learning-based framework for vehicle tracking and visual speed estimation, implemented on the MLK Smart Corridor. Pretrained YOLOV7 was used for vehicles detection. Furthermore, we enhanced the DeepSort tracking model by replacing its feature extractor with a Siamese network trained on the UADETRAC dataset. This modification resulted in a 71% decrease in the trajectories IDSW rate when evaluated against the KITTI benchmark. Our framework can also estimate vehicle speeds using reference object scaling and perspective transformation, achieving a commendable error rate of 0.516 mph. Recognizing the potential of our speed estimator, we identified the need to build a new dataset for visual speed estimation. For this task, a vehicle equipped with an OBU is used to collect speed ground-truth data, and videos were captured simultaneously and passed to the framework for estimating speed readings to be validated. The final dataset contains a total of 1326 annotated instances stamped with their speed readings and grouped into 78 trajectories. This dataset is intended to be used as a benchmark to facilitate research in speed estimation from visual data obtained from urban testbeds.