Abstract:

Convolutional neural networks （CNN） are the key modules for extracting traffic features in deep learning algorithms for traffic state estimation. However， their computational instability on sparse data and multi-modal traffic states limits the accuracy of deep learning for state estimation. To improve the accuracy of CNN-based traffic feature analysis， this paper proposed an encode-decode traffic adaptive convolutional network. Firstly， the proposed network constructed a traffic feature encoding CNN， which uses down-sampling operations to aggregate neighboring traffic information and extracts effective traffic features from sparse data. Secondly， a traffic state adaptive reconstruction CNN is constructed to utilize the extracted features to accurately reconstruct different patterns of traffic states. To represent diverse spatio-temporal structure of traffic states， this CNN introduced prior knowledge to guide the deformation of convolutional kernels. Finally， the performance of algorithm in traffic state estimation under sparse data and different scenarios was validated using taxi GPS data in Changchun City. Experimental results show that compared with advanced algorithms such as LSTM and GAN， the improved CNN alogorithm in this paper improves the estimation accuracy by 6.05 km/h RMSE. Besides， the proposed algorithm has an accuracy difference of only 3.34% RMSE in different traffic scenarios， can provide robust traffic festure analysis support for deep learning-based traffic state estimation.

Key words: traffic management engineering, traffic state estimation, sparse data, convolutional neural network, encode-decode structure