Missing traffic data problem has a significant negative impact for data-driven applications in Intelligent Transportation Systems (ITS). However, existing models mainly focus on the imputation results under Missing Completely At Random (MCAR) task, and there is a considerable difference between MCAR with the situation encountered in real life. Furthermore, some existing state-of-the-art models can be vulnerable when dealing with other imputation tasks like block miss imputation. In this paper, we propose a novel deep learning model TINet for missing traffic data imputation problems. TINet uses the self-attention mechanism to dynamically adjust the weight for each entries in the input data. This architecture effectively avoids the limitation of the Fully Connected Network (FCN). Furthermore, TINet uses multi-dimensional embedding for representing data's spatial-temporal positional information, which alleviates the computation and memory requirements of attention-based model for multi-dimentional data. We evaluate TINet with other baselines on two real-world datasets. Different from the previous work that only employs MCAR for testing, our experiment also tested the performance of models on the Block Miss At Random (BMAR) tasks. The results show that TINet outperforms baseline imputation models for both MCAR and BMAR tasks with different missing rates.