Research On Road Vehicle Information Extraction Method Based On Video Analysis

The vehicle on the road is the key research object in the intelligent transportation system.The acquisition of the vehicle information based on the video can effectively assist the traffic department to supervise and improve the operation efficiency of the traffic system,provide the public with a more efficient and safe travel environment.At present,although there are a lot of research results on the road vehicle monitoring,most of the system is a single function,the requirements to the cameras are inconsistent,and the extraction accuracy is not high enough.In order to solve the following problems,the main work of this paper is as follows:To the current phenomenon on vehicle monitoring system,this paper uses a class of cameras that had been installed on the road to collect video,analyzes the cross-intersection scene under the camera and designs the system frame of extracting the vehicle's individual characteristics and the behavioral characteristics of the vehicle,which makes the traffic statistics,vehicle speed measurement,vehicle mark recognition and queuing congestion detection work together to save resources,and then,compares the basic methods of image processing and machine learning that need to be used when the function is realized.Adaboost combined with SVM is improved and used to locate the license plate in real time;The time and space context visual tracking algorithm is selected to track the license plate accurately and fast;According to V=S/T,sub-calculation of the trajectory and multiplied by different proportion of the length of the track is converted to the actual distance to achieve vehicle speed.The multi-scale parallel convolution neural network is designed to classify the vehicle-logo,and the recognition accuracy is achieved under the condition of less training data set compared with other methods;The existing vehicle-logo positioning algorithm is optimize to adapt to the scene.The algorithm of combining with the background difference method and SVM classifier is designed to achieve queuing congestion detection.Firstly,the scene is analyzed and the virtual detection window is designed.Then,determine whether each detection window has vehicles and foregrounds,and finally the ratio of the foreground and whether there is a vehicle are combined to determine the overall queue length and overflow situation.In this paper,the characteristics of the individual and the behavior of the vehicle are extracted in the same view of the same camera.The article tests the algorithms of each function and verified its accuracy and robustness.The high-level semantic information of the video are showed to the users by the system.