Research On Target Recognition Method For Specific Application Based On GPU

Target recognition technology is an important part in the field of vision,and two kinds of algorithms which are important in recognition technology are the matching algorithm based on gray scale and the matching algorithm based on feature.The principle of the first algorithm is to achieve matching recognition by comparing the distinction of pixel gray value in two pictures,and it has higher precision in the case that the gray level is not much changed,but the matching effect is very unsatisfactory when the angle of view,gray scale and structure are changed greatly.The feature matching recognition algorithm has favorable characteristics in these aspects,so it becomes the hotspot in the field of image matching and recognition.SIFT algorithm is a classical feature matching recognition algorithm,which has invariance of image rotation,scale change and illumination brightness change,and it is popular in scientific research and practical application.Although the SIFT algorithm has many excellent features,it still has some defects in the accuracy and efficiency,so it shows the shortcomings of the algorithm in the environment with high requirements of both accuracy and real time performance.To improve the deficiencies of the algorithm,this thesis proposes an improved scheme.In order to improve the correctness rate,this thesis combines the RANSAC algorithm with the SIFT algorithm,collects some data from the feature point data by RANSAC algorithm after the feature point matching,establishes the parameterization model which accords with all the data,and then can remove most of the error matching points and improve the correctness rate by establishing the model.Aiming at the low efficiency of SIFT algorithm,this thesis analyzes each step in the algorithm principle,carries on parallel acceleration through studying the function of various memory in GPU and the characteristics of itself,and analyzes the parallel programming mode of CUDA platform,designs SIFT algorithm in parallelism,combines the characteristics of memory to divide the algorithm into reasonable parallel tasks,analyzes the parallelism of each step in the algorithm,and completes the parallel acceleration of the algorithm on the CUDA platform,improves the execution speed of the algorithm,and improves the efficiency of the whole system.The experimental platform verifies that the correct rate of SIFT algorithm combined with the RANSAC algorithm is significantly improved compared with the original algorithm,and the correct rate is basically guaranteed to be around 95% in diverse image sizes.After testing the performance of the GPU-accelerated algorithm in distinct scenarios,the results show that the algorithm has no effect on the favorable characteristics of the algorithm itself,and the acceleration effect is more obvious.With the increase of the image quality,the acceleration effect is increased,and then the correctness and efficiency of the algorithm are improved.