Research And Implementation Of Implementation Of Parallel Algorithm For Vehicle Image Retrieval Based On Multi Core CPU And Multi Core GPU

With the increasing number of motor vehicles to use, it's becoming more and more difficulty to find vehicle crime from monitoring image, so how to find the vehicle crime fast from the mass video data is a key question. In this paper, we based on multi-core CPU and many core GPU for monitoring the image region of interest detection algorithm, feature point extraction algorithm, feature matching algorithm and the detection of a series of parallel acceleration to improve the real-time.Firstly, as a large of non interest regions in the monitoring image, and based on image segmentation of region of interest extraction efficiency of the implementation of the algorithm relatively inefficient problem, this paper presents a based on pthreads CPU and multi thread parallel of region of interest extraction algorithm, support for hyper threading in 12 core server, through experimental comparison of 10000 images in image number of serial and parallel execution time validation, relative to the serial parallel algorithm can achieve 13.1 times accelerationSecondly, we consider the problem of vehicle image feature extraction algorithm based on SIFT, the SIFT algorithm in order to improve the image feature robustness to scale changes, rotation, etc., leading to complex calculation process in the algorithm, for tens of thousands of pieces of image data, the execution time can be up to ten minutes or even a few minutes. In order to sift algorithm to accelerate the execution efficiency of large amount of data monitoring images, this paper proposes a four GPU parallel feature extraction algorithm based on, the parallel algorithm through the segmentation of the image data to improve feature extraction efficiency. By comparing the experimental data in 10000 monitoring image data algorithm execution, with respect to a single SIFT GPU algorithm, the parallel algorithm can improve the acceleration of about 3.8 times.Then, in the vehicle image file feature based, because each feature is composed of 128 bit special clinic descriptors and the coordinate of the position. In this paper, we consider according to the Euclidean distance to calculate the vector similarity method to verify is similar to the target image and the reference image. In order to reduce the error matching, the RANSAC algorithm is adopted to eliminate the error matching feature points. On this basis, in order to speed up the matching efficiency, this paper designs and implements a parallel matching algorithm based on CPU.Finally, in the process of finding the whole crime vehicle, the whole algorithm is optimized according to the real environment. In order to make full use of the SIFT features of multi GPU to extract the remaining CPU threads in the execution of parallel algorithms, a parallel algorithm is proposed which combines CPU feature matching with multi GPU feature extraction and parallel execution. The algorithm using GPU feature extraction in immediately with CPU and multi thread parallel execution of matching algorithm, in order to hide the matching algorithm execution time, guarantee the parallel image feature extraction and matching optimization of the execution time of the algorithm, improve the overall parallel execution efficiency.