A monocular vision based three dimensional (3D) dense scene reconstruction technique is presented to achieve fast and accurate 3D stereoscopic modeling in the real environment. This proposed approach localizes accurately a free moving camera in the framework of parallel tracking and mapping (PTAM) algorithm. Based on this self-localization, a variational depth map estimation model is established by using a bundle of image around the keyframe. Discrete depth space sampling strategy is proposed to initialize the variational depth map model and primal dual algorithm is presented to optimize the model afterward. Subsequently, the final 3D scene model can be estimated by integrating the projective camera imaging model. Under the compute unified device architecture (CUDA), the algorithm is optimized in parallel mode by using the graphic processing unit (GPU) hardware, and its real-time performance is significantly improved. The experimental results conducted in realistic scenario demonstrats the feasibility and effectiveness of the proposed algorithm.