Research On Vehicle Automatic Classification Based On Earth Mover's Distance To Dis-similarity Metric

Vehicle automatic classification is a important part of Intelligent Transportation System,which is developed in recent years,it is one of the focused researches on application of image process and pattern recognition.In this paper, current internal and external technique of vehicle classification are synthetically analyzed,especially the study on video image vehicle classification technique. I put forward a new method of vehicle classification based on earth mover's dis-similarity distance metric. The system can gather real time traffic vehicle movement image using video equipment without field demarcation , process the sequences of images using software algorithm and classify the moving vehicles in real time . In this paper we firstly analyze the main characteristic and relevant techniques of vehicle recognition in theory, then discuss the composition of the system and application modal of vehicle classification in system engineering, finally give a simulation realization of vehicle classification system on a practical roadway in system application's realization .The paper emphasizes the following key problems in video image vehicle classification system:1)In the implementation of image vehicle classification system, a new frame for vehicle inspecting is given on a practical roadway, we can easily implement the orientation of vehicle ,and analyze the process effect to variant situation using this frame .2)In the process of vehicle image ,considering the system demand,we adopt several new processing methods: background building method in complex circumstance;dynamic background refreshing;target partition method using L*a*b color space with the intention to eliminate influence of shadow; We use a new dis-similarity metric----earth mover's distance to calculate the dis-similarity distance between the vehicle target and variant classes ,which lead to performance similar to human feeling .3) we use image shape local context descriptor to describe the image target in order to avoid the dis-similarity distance being affected on image vehicle translation and scale variance ,and give the experimental data and analysis of the result to testify the invariance to translation and scale.4) Shape matching problem belongs to the high dimension classification problem ,its computing complication is high , in order to resolve the intime response of real time system ,we use a newly approximate algorithm---metric embedding ,in order to split the difference between the precision and real-time and give the experiment result .We testify the simulation system using the practical roadway vehicle information , and achieve the anticipated design effect.