An improved tucker decomposition-based imputation method for recovering lane-level missing values in traffic data

High-quality lane-scale traffic data is of great importance to the intelligent transportation system. However, missing values are sometimes inevitable due to the failure of the detectors or the low penetration rates of the connected automated vehicles. Most existing data recovery methods concentrate on traffic data of a whole section and ignore the spatio-temporal correlations between lanes. To this end, this paper organizes the lane-scale traffic data into tensor patterns that can simultaneously consider the spatio-temporal dependencies of traffic flow. Then an improved Tucker decomposition-based imputation method (ITDI) is proposed to recover the missing values of the traffic data by extending the Tucker decomposition model with an adaptive rank calculation algorithm and improved objective function. Using the real-world traffic data to construct multiple datasets with three missing scenarios and different missing rates, the performance of the proposed model is evaluated and compared with that of state-of-the-art data imputation methods. The experimental results indicate that the ITDI method has better performance than the baseline models in terms of imputation accuracy. Besides, the ITDI model can adapt to typical missing scenarios and keep stable under different missing rates.