Research On Video Detection Algorithm For Vehicles Red-light Running Based On GPU

With the number of vehicles increasing, traffic pressure is increasing dramatically, androad traffic accidents caused by vehicle violations are also increasing, so many countries inthe world have strengthened the research on intelligent transportation system and itsapplication. Vehicles red-light running detection is an important part of intelligent trafficsystem, because it is important for reducing traffic accidents, regulating traffic order ofintersections and ensuring the smooth flow of traffic. Vehicles red-light running detectionbased on video has many merits, such as wide monitoring range, low cost, and so on, so it hasbeen deeply studied and widely used in recent years. Moving object (vehicles) detectionalgorithm based on Gaussian mixture model is the most popular algorithm in this area, butwhen this algorithm is used for vehicles red-light running detection, there are severalproblems to be solved, for example, unreasonable initialization parameters and poorinstantaneity. To solve these problems, this paper studies parameter initialization in Gaussianmixture model, GPU parallel computing, and so on. The main work and contributions are asfollows:(1) In real surveillance video, if the first frame collected during a red light is used toinitialize Gaussian mixture model parameters, then vehicles in the first frame usually makethe Gaussian mixture model initialization parameters unreasonable. This paper presents animproved algorithm for Gaussian mixture model initialization, which constructs backgroundimage using the video data during the first red light, and then uses the background image toinitialize Gaussian mixture model parameters. Experimental results show that improvedalgorithm proposed in this paper improves the accuracy of vehicles red-light runningdetection.(2) Because the heavy computation of vehicles red-light running detection algorithm, thispaper designs and implements vehicles red-light running detection algorithm with GPUparallel computing to meet the requirement of real-time detection. Experiment results showthat GPU-based parallel algorithm proposed in this paper can effectively improve theinstantaneity of the vehicles red-light running detection algorithm.