Towards a parallelization and performance optimization of Viola and Jones algorithm in heterogeneous CPU-GPU mobile system

Parallel computing on heterogeneous multiprocessor architecture is a new technique used to tackle the complexity of actual media applications. Such technique is used on an embedded architecture composed of 2ARMs coupled to a GPU. In this paper, an approach for real time Viola and Jones face detection algorithm using CPU-GPU based platform is presented. First, the application is implemented and parallelized on two identical ARM CortexA9 CPUs using tasks and data levels parallelism. This technique does not achieve the timing objectives. To ensure greater performance while reducing energy ratio, we extend our parallelization technique to support a GPU as an accelerator to perform non graphical tasks. OpenCL, the heterogeneous parallel programming model, is used to ensure communication between CPU and GPU.