Research On Vihicle Flow Detection Based On Video Image And GPU Acceleration

With the increasing flow of traffic,traffic problems come along,so the development of Intelligent Transportation Systems(ITS)has become crucial.It is an important part of the real-time detection,analysis and processing of traffic information.The research on video image and vehicle traffic statistics is an important topic in the research of intelligent transportation system.This paper analyzes the video images based on video images and GPU-accelerated vehicle traffic detection methods and statistics of automobile traffic,and proposes the AdaBoost classifier to discriminate video vehicle detection methods and match feature statistics based on the target vehicle's center of gravity as the target.Based on the above,we accelerate the processing of vehicle traffic statistics algorithm by GPU,and achieve real-time detection.The main contents of this paper are as follows:(1)Vehicle target detection,the past and widely used methods are: color and pattern based methods,template matching,spot tracking,contour tracking,background subtraction,Gaussian mixture model,and optical flow.However,these methods have drawbacks in terms of efficiency or implementation cost.This article differs greatly in the type and color of different vehicles,but the characteristics of the license plate area are similar.Using this feature,a vehicle detection method based on license plate information,through off-line learning,training detects the license plate area Adaboost classifier.First,the surveillance video vehicle automatic detection method extracts the area of interest and performs motion modeling;then,it detects the license plate area target by sliding.Improve the classifiers in the area of interest and use the clustering method to merge the location information of the target vehicle.(2)In the modeling of sports background,this paper introduces several modeling methods.The hybrid Gaussian modeling method has a good detection effect.The disadvantage is that it has a large amount of computation and it is difficult to achieve real-time performance.The article uses GPU platform CUDA compilation environment to design and parallelize the Gaussian modeling algorithm.Experiments show that the parallelization of the algorithm is faster than the CUDA operation speed in OpenCV,and can be about 6 times faster than the serial algorithm.The acceleration effect is obvious.(3)Traffic statistics and counting,analysis improves the feature extraction of moving target vehicles.After feature extraction,the image is binarized,the target vehicle information and data flow are obtained,and then the two-bit boundary of the target area is modeled.Select the matching criteria that is suitable for detecting the target area and count the traffic volume.The algorithm in this paper not only has accurate detection target and high detection efficiency,but also can realize cross lane counting and multi-lane counting.The measured data show that the algorithm of this paper can accurately detect the traffic volume,and over time,it can also calculate the vehicle flow data over a period of time.The statistical effect is stable,and the actual measurement accuracy is over 95%.The experimental results of the above methods have proved this paper The algorithm has satisfactory performance in terms of efficiency and implementation cost.