Parallel Face Detection CUDA Accelerating

As an important issue in computer vision and pattern recognition, face detection hasbeen used widely in information security, video surveillance and biometrics, etc. The Ad-aBoost algorithm based face detection method proposed by Viola et al in2001is the firstand one of the most popular techniques for real-time face detection.In the present thesis, we parallelize the AdaBoost based face detection method andimplement the parallelized method on NVIDA GPU by CUDA.In the framework of the presented parallelized face detection method, the first step isconverting the input color image to gray model, then the gray image is resized to a pyramidimage at different scales so that all size faces in the original image can be detected efficientlywhen using fixed-size scanning windows. Both color space converting and image resizingare implemented on GPU with parallelized method, which is hundreds of times faster thanthe CPU implementation.The next step is creating integral image and squared integral image of the pyramidimage on GPU. The parallel prefix sum algorithm and image transposing method are adoptedso that the integral image and squared integral image can be computed efficiently. Withcompared to the CPU-based integral image computation, our method is much faster.When detecting faces with the integral image and squared integral image, the kernelson GPU scan and evaluate all sub-windows in parallel. In this step, the kernels can be accel-erated by storing the classifiers in the constant memory and mapping the integral image tothe shared memory on GPU. We also designed a efficient method to processing the detectingresult using a bit array.The experiment results showed that the parallel face detection proposed in this thesisis more efficient than the serial face detection method, and could achieve an over30xspeed-up compared with the CPU implementation. Moreover, our method could achievereal-time face detection for1080P HD videos.