Unpaired Learning of Roadway-Level Traffic Paths from Trajectories

Weixing Jia, Guiling Wang, Xuankai Yang, Fengquan Zhang

Abstract: Traffic path data can be used as the basis for traffic monitoring and other technologies, which is essential for developing traffic-related technologies. Traditional methods of traffic path data extraction can no longer meet the needs because they cannot solve the problem of lacking standard benchmark data that may exist in the traffic field. Deep learning-based path extraction methods using large-scale data are a class of promising approaches. However, most of the deep learning-based path extraction methods are supervised and rely on paired training data. This paper proposes an unpaired learning method for fine-grained roadway-level paths from trajectory data based on CycleGAN. The method constructs spatio-temporal features based on HSV color space from trajectories which can enhance the model’s ability to recognize the roadway details. It transforms the features using convolutional layers, which can preserve the spatio-temporal information of the features, thus making the extraction results more accurate. We conduct experiments using urban and maritime traffic trajectory data and compare the proposed method with the state-of-the-art methods. The results of our model have more roadway level details, higher precision and F1 score than the other existing unsupervised traffic path learning methods.

External IDs:dblp:conf/colcom/JiaWYZ21