GPU-based real time motion tracking in high definition videos

We propose in this work to exploit parallel (GPU) and heterogeneous (Multi-CPU/Multi-GPU) architectures for boosting performance of image and video processing algorithms. Our interest for accelerating these methods is due mainly to the high increase of calculation intensity of such applications, and the strong growth in the size of multimedia objects (HD/Full HD) during the recent years. We propose in this work a model for intensive multimedia (single image, multiple images, multiple videos, video in real time) processing based on exploiting the full computing power of machines (Multi-CPU/Multi-GPU). This model enables to select firstly the computing units (CPU or/and GPU) for processing, and secondly the methods to be applied depending on the type of media to process and the algorithm complexity. The proposed model exploits efficient scheduling strategies which allow a full exploitation of hybrid machines. Moreover, it enables to reduce significantly data transfer times thanks to the efficient exploitation of GPU memories and to the overlapping of data copies by kernels executions on GPU. Indeed, we have implemented several algorithms on GPU such as optimal contours extraction based on Deriche-Canny method, points of interest detection, silhouette detection and optical flow computation usedfor motion estimation. Experimental results have been obtained by applying the proposed model on different types of media (HD images, databases of medical images, HD videos). These results showed a global speedup ranging from 10 to 50, by comparison with sequential CPU implementations.