GPGPU-Based Parallel Computing of Viola and Jones Eyes Detection Algorithm to Drive an Intelligent Wheelchair

The emergence of low-cost perception sensors and the abundance of inexpensive machine vision cameras, allow incorporating intelligent human user interfaces into electrical powered wheelchair platforms. Currently, these interfaces rely on computer vision and image processing algorithms. However, these algorithms require significant computing power for the autonomous navigation. A new trend to satisfy this constraint is the use of a GPU in cooperation with a CPU to accelerate computer vision and massively parallel and time-consuming applications. In the present paper, we propose a GPU-accelerated method to parallelize the well-known Viola and Jones based eyes detection algorithm. The implemented algorithm was applied to a human machine interface with a rapid feedback reaction to control a smart wheelchair designed for people with disabilities. We focused on the implementation of the system on a mobile high-performance computing CPU/GPU platform. The added value of the proposed implementation is the parallel strategies that we adopted for the Viola and Jones algorithm using OpenCL API. Compared to a multi-threaded version, for a 640 * 480 pixels image, our OpenCL C-based GPU version took around 55 milliseconds instead of 116.791 milliseconds for the multi-threaded version. By achieving nearly 2.1 speedup without precision degradation, the proposed heterogeneous strategy is useful for implementation of other computer vision methods.