Road Network Similarity-Based Transfer Learning Method for Traffic Volume Estimation in Undetected Road Segments

Estimating traffic state, particularly traffic volume, is crucial in Intelligent Transportation Systems (ITS). Due to the absence or malfunction of detectors, some road segments are undetected, leading to a complete absence of volume data and thereby weakening the traffic monitoring capability of ITS. The existing estimation methods are either inapplicable to this scenario or yield poor results due to a lack of available data, which will compromise the traffic monitoring capability of ITS. To handle it, this work proposes a novel Road Network Similarity-based Transfer Learning method (RNS-TL) for real-time traffic estimation. Firstly, the Small-scale Road Network Similarity Evaluation Module (SSEM) is initially proposed which aims to identify the most similar road segments and their small-scale road networks for the undetected segments, serving as the source domain for transfer learning. Then, based on SSEM, a transfer learning framework is proposed where a traffic estimation model trained on the source domain is fine-tuned for the target undetected road segment. Finally, the results from two real-world traffic cases show that the estimation errors, MAE and RMSE, for the proposed method are 7.813 and 6.383, and 10.689 and 8.892, respectively, outperforming all comparison methods.