The GPU Implementation Of Pedestrian Detection Algorithms Based Centrist Feature

Pedestrian detection is widely used in smart robots and intelligent transportation.In recent years,the technology has become a hot study in computer vision and pattern recognition.There are many challenges in pedestrian detection algorithm,including real-time and robustness requirements.The CENTRIST feature which describes the critical contour information for the object is more suitable for pedestrian detection than the widely used HOG features.The pedestrian detection algorithm(C4 algorithm)using a cascade classifier and CENTRIST feature than HOG algorithm in detection speed and detection accuracy is more advantage.In this paper,C4 algorithm is parallelled on NVIDIA GPU by CUDA.In the framework of the presented parallelized pedestrian detection method,the first step is converting the input color image to gray model.The gray image is then resized to a pyramid image at different scales.The Sobel image is constructed by using the Sobel operator to filter the original image.The CT image is constructed by calculating the CT value of each pixel of the Sobel image.Pyramid image,Sobel image,and CT image are computed on GPU with parallelized method,which are dozens of times faster than the CPU implementation.In the detection phase,pedestrians are detected by using linear SVM.The auxiliary image and integral image are created.The classification results are calculated by using integral image.HIK SVM can achieve higher classification accuracy than linear SVM,so further detecting is necessary by HIK SVM.The final classification results are calculated by summing the HIK SVM data indexed by the CENTRIST feature.In the detection phase,The GPU implementation is more than ten times faster than the CPU implementation.The performance of parallel algorithm is improved by applying many optimization methods,such as memory access optimization,data transmission optimization and control flow optimization.The pedestrian detection algorithm in this paper achieves 108 fps which is 10 fps faster than the existing fastest algorithm and achieves real-time detection for large-size images.The parallel pedestrian detection algorithm can achieve speedups of 17 x over the serial algorithm while without compromising any accur acy.In addition,the pedestrian detection algorithm in this paper is a general algorithm that can be extended to the detection of animals,plants and other objects.