Towards an End-to-End Framework of CCTV-Based Urban Traffic Volume Detection and Prediction

Near real-time urban traffic analysis and prediction are paramount for effective intelligent transport systems. Whilst there is a plethora of research on advanced approaches to study traffic recently, only one-third of them has focused on urban arterials. A ready-to-use framework to support decision making in local traffic bureaus using largely available IoT sensors, especially CCTV, is yet to be developed. This study presents an end-to-end urban traffic volume detection and prediction framework using CCTV image series. The framework incorporates a novel Faster R-CNN to generate vehicle counts and quantify traffic conditions. Then it investigates the performance of a statistical-based model (SARIMAX), a machine learning (random forest; RF) and a deep learning (LSTM) model to predict traffic volume 30 min in the future. Tests at six locations with varying traffic conditions under different lengths of past time series are used to train the prediction models. RF and LSTM provided the most accurate predictions, with RF being faster than LSTM. The developed framework has been successfully applied to fill data gaps under adverse weather conditions when data are missing. It can be potentially implemented in near real time at any CCTV location and integrated into an online visualization platform.