The Research Of Vehicle Flow Detection Technology Based On Video

In recent years, video-based vehicle flow detection technology is widely used because ofits flexible parameter settings, wide detection range, easy installation and maintenance.According to the traffic videos from camera as the research object, this paper focused onvehicle detection and vehicle tracking algorithms of vehicle flow detection technology. And aprototype of all-weather vehicle flow detection system was developed based on VC++andOpenCV. The main works of this paper are as follows.First, the collection method of video samples was researched. To set up a video samplelibrary, several traffic videos in different times, different weather and different trafficconditions were shot with a household DV. The library was used for algorithm tests andexperiments.Second, an improved method of Gaussian Mixture Model was proposed. For the defectof long-term “ghosts” in Gaussian Mixture Model, a new update method of weights andvariances was employed to accelerate the elimination of “ghosts”. In Gaussian Mixture Model,all of the pixels are modeled by fixed number of distributions, thus main memories are wasted.To save memories, a self-adapting method was adopted. For the pixels whose distributionnumbers are not up to maximum, they are modeled by variational number of distributions toeffectively decrease the total number of distributions. The results indicate that the improvedGaussian Mixture Model has great advantages in the elimination of “ghosts” and computingspeed. On the basis of the Gaussian Mixture Model, a shadow detection algorithm based oncolor and texture was proposed. In order to distinguish shadow pixels and non-shadow pixels,the fact is used that the brightness of pixels in shadow region is lower than that in background,while the colors and textures remain the same. By computing and comparing the colordistortion, brightness distortion and first-order gradient，the shadow pixels are recognized.The experimental results show that the algorithm can accurately detect the shadow pixels.Third, the vehicle tracking algorithm based on bounding rectangle and Kalman Filterwas researched. Two matching indicators, area index and distance index, were proposed byusing the characteristic that the area and moving distance of the same vehicle in two adjacentframes are close. Combined with Kalman Filter, the tracking was complete.Fourth, vehicle flow detection algorithm based on headlights at night was studied. Theheadlights were extracted with region growing algorithm whose seed points were segmentedby self-adapting threshold process based on gray histogram. The vehicle flow and vehicletrajectories were given by headlights pairing and tracking. Finally, a prototype of all-weather vehicle flow detection system was designed and itsmain function modules were introduced. To get the vehicle flow data, experiments were donewith the videos in the library. By comparing the results with those obtained by the virtual coilmethod, it proved that the system has a higher accuracy of vehicle flow detection.