Marker-Less 3D Human Motion Capture in Real-Time Using Particle Swarm Optimization with GPU-Accelerated Fitness Function

Reconstruction of 3D human motion in real-time using particle swarm optimization with GPU-accelerated fitness function

“…Using CUDA we turn the GPU into a powerful parallel processor, whereas using OpenGL we utilize the same GPU hardware to render the required images. Because CUDA and OpenGL both run on GPU and share data through common memory, the CUDA-OpenGL interoperability is very fast [25].…”

“…That means that we effectively utilize the rendering power of OpenGL to render the 3D models in the requested poses, whereas the CUDA threads match such rendered models with the image features and perform the tracking. Because CUDA and OpenGL both run on GPU and share data through common memory, the CUDA-OpenGL interoperability is very fast in practice [25,41]. In the motion-tracking methods relying on matching the projected models with the camera images, the most significant computational overheads are associated with the rendering of the 3D models [23,25].…”

“…Because CUDA and OpenGL both run on GPU and share data through common memory, the CUDA-OpenGL interoperability is very fast in practice [25,41]. In the motion-tracking methods relying on matching the projected models with the camera images, the most significant computational overheads are associated with the rendering of the 3D models [23,25]. Thus, the OpenGLbased 3D model rendering can shorten the whole motiontracking time considerably.…”

“…Because CUDA and OpenGL both run on GPU and share data through common memory, the CUDA-OpenGL interoperability is very fast in practice. As indicated in [25], time needed for mapping the rendering result is equal to 0.4 ms, and the overall execution time required for determining the objective function values in the single iteration of PSO is equal to 4.4 ms, for a swarm consisting of 96 particles and measurements done for LeeWalk sequence.…”

“…We demonstrated experimentally that on P1S image sequence, which is used in this work, we can achieve fivefold speed-up, and full-body tracking with the average error equal to 50.6 mm. In [25], we demonstrated that in 3D-model-based human pose tracking a considerable shortening time of calculation of the fitness function can be achieved through the use of OpenGL-based rendering of the model in the requested poses. We measured the execution times of the main components of the fitness function using CPU, CUDA, CPU-OpenGL and CUDA-OpenGL.…”