Research Of Traffic Scene Understanding Algorithm Based On Conditional Random Fields

Safety is the eternal theme of transportation industry.In recent years,with the rapid growth of car ownership,traffic accidents occur more and more frequently,the harm it gives the human society is serious.In this context,car safety auxiliary driving technology is paid more and more attention.As an effective means to reduce traffic accidents and reduce accident loss,it become the research frontier in the field of traffic engineering,represents the development trend of future vehicles.With the in-depth development of automotive safety auxiliary driving research both at domestic and overseas,part of the relevant technology was gradually mature,and began to go to the transition,but there are still quite a lot of technical problems remain to be further solved.Road scene understanding is an important part of the auxiliary driving safety and becomes a focus to research.Based on the analysis of the basis of summarizing the domestic and foreign various scenarios to understand detection method,this paper puts forward a more effective scene understanding detection algorithm.We study the road scene understanding algorithm based on the road scene image semantic segmentation.The road scene can be divided into three categories:the sky,road and obstacles.Due to Conditional Random field(CRF)Model can merge characteristics and context information together using many types of images,scene understanding based on Conditional Random field technology is the research hotspot in the field of computer vision in recent years.The core idea of this method is to construct a probability graphical models(PGM)with the observed data,gives a potential function corresponding to the image structure,according to the learning and inference methods of Conditional Random field to get semantic labels with the highest probability as a sign of the final results.In order to more accurately and quickly identify the scene image categories.Conditional Random Field theory(Conditional Random Fields,CRF)was applied to do identification mark of scene area.Respectively establish the CRF model based on regular pixel block and the CRF model based on super pixel block In the process of building model,this paper mainly do the following work innovation:1.The introduction of the Support Vector Machine(SVM)classifier and Local Variation(LV)segmentation algorithm technology.2.Makes the model has a better effect using soft max regression method and maximum likelihood estimation of piecewise training,the quasi Newton iterative method(Limited Broyden Fletcher Goldfarb Shanno,LBFGS)is used to study the parameters of iterative optimization and used the average Field training algorithm(Mean Field)to infer the model.3.Multiple CRF model is established based on the multiple segmentation to achieve higher recognition rate of road scene image semantic segmentation.This algorithm was tested in public data sets ICCV09DATA,GC(Geometric Context)and real scene in the video,the recognition rate are 92.12%and 92.00%respectively.