Parallel Implementation of 3D Model Reconstruction of Monocular Video Frames in a Dynamic Environment.

Real-time three-dimensional (3D) reconstruction has been widely explored in several domains in motion capture, robot navigation, augmented reality, and autonomous driving. It is considered one of the most efficient solutions to overcome several problems, such as occlusion and collision in computer vision and computer graphics. This process is computationally intensive and it considers a bottleneck for real-time realistic interaction applications which need a quick response to achieve action in real-time. The Monocular 3D model reconstruction (M3DMR) is considered one of the possible solutions to build an accurate 3D reconstruction in a dynamic environment. However, the proposed framework needs a high computational power to create a single frame. Graphics processors unit (GPU) architecture is used to improve the computational time of M3DMR. This study discussed how to maximize the benefits of using GPU resources by using several optimization techniques that enable GPU architecture to achieve the best possible performance for the M3DMR. Different multicore heterogeneous systems are used to evaluate the performance of the proposed framework. Experimental results confirm that our parallel implementation of 3D Model Reconstruction of Monocular Video Frames is valid for different GPU architectures. The proposed parallel implementation can execute 28 FBS compared with the serial version that executes one frame in 30 min. [ABSTRACT FROM AUTHOR]