Real-time pedestrian detectionl Using OpenCL
Rong Sun, Xuzhi Wang, Xuannan Ye · 2014
Pedestrian detection is a challenging task, due to wide variety of appearances, especially in complex real world scenes. The use of real-time pedestrian detection is of great use for a broad range of applications in multiple domains, such as surveillance and Intelligent Transportation System. In this paper we present a fast implementation of a robust pedestrian detector by using OpenCL, which is a novel open standard for heterogeneous computing. OpenCL allows for scalability to better performance and different types of hardware, with minimal changes to the implementation. To show the portable ability and performance of the new implementation of algorithm based on HOG, the algorithms are executed on the three different platforms, including CPU + NVidia GPU and CPU + AMD GPU heterogeneous system. By using a GPU as execution device, we exploit the data parallelism opportunities of the algorithm. In pedestrian detection, HOG is a very good algorithm but long running time because of its complexity. In this paper, the HOG and SVM algorithms will be optimized with OpenCL technology to achieve the goal of real-time requirements. On a single CPU + GPU machine, we reach 36 fps on the premise of algorithm's portability.