Research Of Object Detection Algorithm Based On CUDA Acceleration

Object detection is the basic subject of image processing.It is the basis of identification,tracking and classification tasks.Object detection have been widely used in military and civil fields.With the increase of the amount of data and the complexity of detection algorithm,detection software based on CPU serial implementation has been unable to meet the real-time requirements detection tasks.In recent years,the GPU common computing technology can provide powerful parallel computing power,which provides the possibility and research direction for the real-time realization for the task of object detection.Therefore,the research of object detection algorithm based on GPU is of great significance to improve the real-time performance for object detection task.This thesis gives a brief introduction to GPU and CUDA programming model,discussed the performance of the existing object detection algorithm,and selected HOGSVM detection algorithm based on machine learning and the algorithm using convolutional neural networks for object detection based on deep learning as the research object.Then detailed introduces the theory of the algorithm,discuss the parallelism and CUDA parallel implementation of the algorithm,and optimization.Finally,the process of using the algorithm to carry out the actual object detection task is further parallelized and optimized to maximize the acceleration ratio.The detail work of the thesis is as follows:(1)Researching the CUDA programming model,storage model,CUDAprogramming and optimization process,Discussing the applicability,accuracy and real-time on existing object detection algorithm.HOG-SVM algorithm and object detection algorithm using convolutional neural network,which has good detection performance and poor real-time performance,is selected as the research object.(2)Researching the theory and detection process of HOG-SVM algorithm,analyzing the parallelism of the proposed algorithm.The HOG-SVM algorithm is implemented in parallel and the optimization is carried out.After that,the process of object detection using HOG-SVM algorithm is further optimized to make the whole detection process reach 20-80 times acceleration ratio.(3)Researching the theory of convolutional neural network,building the basic convolutional neural network.On this basis,a network model can be built to accomplish the object detection tasks.Then analyze the parallelism of convolutional neural network training and detection algorithm.Forward process,backward process and weight updating process of the convolution neural network are implemented in parallel through CUDA,and then optimize the algorithm.Finally,the process of multi sample batch training and target detection is further parallel implemented and optimized,making the training and testing process of convolutional neural network reach 40-50 times acceleration ration.