Recently rapid and efficient localization of lung tumors based on single cone beam CT (CBCT) projection has attracted great interest in the radiation oncology community. High quality 4D-CBCT benefits accurate monitoring of moving targets during radiation therapy (RT). However, conventional iterative 4D-CBCT reconstruction process is heavily time-consuming. To address these issues, a motion sensitive cascade model was proposed. The proposed cascade model is composed of dual attention mechanism together with residual network (DA-ResNet) and a principal component analysis (PCA) model. It maps single projection from different breathing phase to each phase 3D-CBCT for achieving real time 4D-CBCT. The dual attention mechanism focuses on the motion information of both low-level features and high-level features to improve accuracy and efficiency of the network. The PCA model ensures a real time motion representation scheme. Compared with state-of-the-art networks, the proposed method outperformed them in the quantification labels of mean absolute error (MAE), R-squared (R2), normalized cross correlation (NCC), and structural similarity index measure (SSIM). This experiment was verified both on simulation and clinical data to support the conclusion.