Research And Implementation Of Pedestrian Detection Algorithm Based On CUDA

With the continuous development of national economy and the increasing of people's living standard,cars have become an important vehicle travel.In China,the amount of individual cars is increasing and the cars also greatly improve the convenience of people's travel,but the traffic accidents are increasing.As an important part of the driving assistant system,the pedestrian detection is very important to ensure the safety of the driver and the safety of the life and property of the pedestrian.But accuracy enough pedestrian detection system needs a lot of calculation,it's time-consuming and difficult to achieve the result of real time.The emergence of GPU and promotion for our parallel speed provides a new way to study algorithm.Calculation of Unified Device Architecture(CUDA,Compute Unified Device Architecture)opens the use powerful GPU computing power to do general door,which allowing developers develop computing power of GPU in a friendly development environment.The new launch of Jetson NVIDIA Pro has strong performance,it can provide a fully functional NVIDIA CUDA platform and make it easy for developers to build and test all kinds of cars and computer vision applications.The main research of this thesis is to solve the contradiction of pedestrian detection algorithm accuracy and the real-time performance,how to transform the pedestrian detection algorithm for parallel and resolve this contradiction with GPU acceleration and experiments it on Jetson Pro.In this thesis,I analysis and study the research status of pedestrian detection algorithm,select Haar-like features described as a pedestrian,use the optical flow computation,forecast future pedestrian location on the basis of detected results and improve algorithm robustness.CUDA programming model is introduced and its software and hardware architecture and NVIDIA Jetson Pro of the performance and characteristics of the development platform,parallel algorithmon this basis.In this thesis,I analysis the time-consuming and characteristicof the serial algorithm according to the computational complexity through the deep study on the pedestrian detection algorithm.Then I select part of integral figure calculation and AdaBoost cascade classifier to parallel implementation based on pyramid optical flow computation algorithm.And combined with the characteristics of CUDA,some optimization strategies are put forward,which makes the algorithm more suitable for CUDA architecture and achieves better execution efficiency.In this thesis,the development platform for experiments is the Jetson Pro NVIDIA vehicle.And then I contrast time consuming both on the serial algorithm and the parallel algorithm.The experimental results show that the parallel pedestrian detection algorithm is significantly improved compared with the serial pedestrian detection algorithm and the real-time performance is improved.