GPS trajectory data have become increasingly useful in traffic analysis and optimization. Nevertheless, due to sampling and communication-related issue, such trajectories suffer from data missing problems, and they further render a low quality of raw data for subsequent research. To address this problem, in this work, we propose a recurrent neural network based encoder-decoder deep learning approach. The head-direction information of trajectory, defined by the radius of curvature, is utilized together with the displacement attributed by an attention mechanism to learn from past trajectory points with different priority. Additionally, a smoothing data post-processor is adopted to make the reconstructed trajectories authentic. To evaluate the performance of the proposed reconstruction approach, a series of comprehensive case studies are conducted, which indicates that the proposed approach significantly outperforms baselines, such as the reduction of the missing impact to the original data and improvement in the prediction accuracy.