Research On Scene Parsing In Intelligent Transportation Surveillance System

Scene parsing system plays an impartant role in Intelligent Transportation System(ITS), which has been widely applied in video compression, video classification and video retrieval of traffic surveillance. Traditional scene parsing methods mainly aim at processing sample nature scenes, which has a high accuracy of segmentation and recognition. However, traditional methods have an imperfect performance at large scale complicated ITS scenes. In this thesis intelligent transportation surveillance scenes are mainly focused and a non-parametric scene parsing system is constructed for it, which is a data-driven method. As the same time, a new method of zebra detection in the traffic scenes is proposed, which has an improved performance in accuracy for the scene parsing. Based on the problems above, the thesis content is as followed.1. In the thesis, large scale scene backgrounds are extracted from the traffic surveillance videos. LabelMe server are constructed and we label the scene background and build a new dataset about traffic surveillance with LabelMe web tool. A baseline is proposed on this traffic scene dataset.2. Traditional scene parsing system is constructed and then we improve the performance of it by applying the multi-chanel and finer granularity GIST global descriptor.3. We propose a new novel and robust approach for the detection and location of zebras, which is based on MSER and extended RANSAC. Promising experiment results have shown the proposed method has a high detection rate and is robust to the different zebra angles and various illuminations.4. Refering to the traditional non-parametric scene parsing system weakness, we improve the performance of parsing system, by fusing the zebra dection and non-parametric method.5. We develop the scene parsing GUI, which can display the parsing results, extract the regon egdes coordinates and save them in database so that other image retrieval system can reach them.