Design And Implement Of Robot Face Detection Based On GPU

In the field of robotics research, face detection, as an important research topic in this area, has made stricter requirements in the detection of real-time performance in the practical environment of the robots. AdaBoost-based face detection algorithm is the most widely used face detection algorithm, which ensures the detection rate and greatly improves the detection speed in the same time.In this paper, the main work is divided into two parts. On the one hand, the work is to conduct the related experiments and choose the best results based on the existing parallel AdaBoost face detection algorithm. At the same time, making the further optimization and the implementation to part of the core algorithm; on the other hand, the work has focused on building the robot’s face detection system, and the whole system is mainly composed of GPU parallel face detection module, human face tracking module, face image sample management module. Finally, combining the above two parts of the work, the hardware equipment based on the nvidia GTX960 and CUDA parallel computing architecture, I successfully complete the construction of the robot’s face detection software system based on GPU acceleration.In the parallel process based on face detection algorithm of Adaboost, I mainly complete the parallel color space conversion from the original color image to the gray image; for the gray image transformation, I analyze and compare the effect of the image to zoom in with the nearest neighbor interpolation method and the bilinear interpolation method by many experiments. Bilinear interpolation method is finally determined to complete the image zooming calculation; In the process of synthesis of pyramid images, I optimize the optimal solution of the problem about the height selection on left and right during the image stitching process; In the process of calculation in pyramid image integral figure and square integral figure, after comparing the different prefix sum algorithms, I finally choose parallel prefix sum algorithm combining the image blocks thought and the image transpose(to reduce the time consumption of access and storage) to complete the high efficient parallel computing in this part; In the process of detection, I finish the parallel scanning detection about pyramid integral image and square integral image by using the method of parallel scanning window, and output the results.In the process of building the robot’s face detection system based on GPU acceleration, my main work include requirements definition and general conception of the system, and finally achieve and verify the system. The system is divided into face image management module, GPU face detection module and face tracking module. According to the results of the practical application, the robot’s face detection system can effectively and accurately detect the faces in real-time video from the robot, and achieve real-time tracking, which meet the needs of the system.