Research On Identification And Quantization Of Traffic State For Expressway Based On Video Sequences

Intelligent Transportation System(ITS) is a trend for worldwide road trafficwhich the transportation researches are striving for developing universally. It’s acrucial subsystem and function identifying and classifying the traffic stateefficaciously. At present, classification of traffic state generally depends on theadvanced technology of information collection, where videos are preferred attributingto their low-costs and automation. However, traditional methods of classificationbased on traffic parameters do not perfectly adapt to traffic state. The accuracy ofthem is likely decreasing with congestion. Besides, it is incomplete to describe actualcondition of road network if dividing whole traffic state into a few classifications. Socombination of historical data will be helpful for interpreting and predicting trafficstate.In this paper, we propse a method of classifying and quantifying traffic statebased on spatio-temporal information (STI) from video sequences. We fix camera andfocus on the urban expressway. As controlled trial, we have extracted trafficparameters to classify state from undifferencially experimental conditions. This paperhas designed and implemented the following three main algorithms separately: trafficparameters extraction from STI, visual features extraction from STI and classificationand quantification based on the above information. The main contents of this paperare expounded as following:1. In the section of traffic parameters extraction, we firstly capture objects’s realinformation by dividing region of interests and calibrating camera. We can get twokinds of STI by accumulating content of virtual detection and trace lines. This paperproposes a binary method with self adapted threshold to extract vehicles from STIs.Then current algorihms of parameters extraction are optimized by integrating PVI andEPI images. Parameters include occupancy of time, volume and average speed.2. In the section of visual feature extraction, a method is proposed whichextracting features from PVI image directly. At the first step, the image is processedby Principal Component Analysis (PCA) to remove redundant data. Then we willinsert lable about state to PVI after PCA. Finally, Fisher Linear Discriminant Analysis(FLD) is used to further extraction. Experimental results testified the processed imagematrix can be distributed by classes.3. As to the section of classification and quantification, we input both traffic parameters and visual features into Support Vector System (SVM) to train andconstruct classifier. Traffic states are divided into for classes including fluent,non-congestion, congestion and terrible congestion. The method is testified on thedifferent kinds of kernel functions of SVM and other classifiers such as BP, Baysianand RBFNetwork. By comparasion, the algorihm of SVM based visual featureperformed outstanding. Finally, quantitative value can be obtained by computingdistance from data point to adjacent classification superplane.4. In the final section, this paper design and implement MFC Demo of trafficstate classification and quantification based on STI. Systematic functions andimplementation model are illustrated and revealed detailedly with specific videosequences. All algorihms in this paper are embedded in Demo.