Research Of Scene Understanding Algorithm Based On Monocular Vision

With the accelerating pace of urban life, car has become a part of people’s life, but meanwhile, the frequency of traffic accident increases also, which turns into globally widespread concern. Safety Driving Assist can solve the problems of traffic safety. Scene Understanding is essential to solve the traffic safety problem, which is the research focus in various nations. On the base of analyzing and summarizing the detection methods of Scene Understanding abroad and home, this thesis proposes a new method.This thesis carries on study about Scene Understanding, makes image segmentation of it, and divides it into three categories:sky, roads and obstacles. In order to mark the scene image categories more accurately and quickly, Extreme Learning Machine (ELM) is used to identify and mark the scene region. Firstly, to extract the scene region HOG (Histogram of Oriented Gradient) features, apply BW dimension reduction methods, use ELM classifier for classification, train three ELM classifiers offline (sky, land, obstacles), and then, identify the corresponding scene categories. Although the scene has been made semantic annotation segmentation, the test results are not satisfactory; therefore, further consideration has been given to other image information, to classify semantic segmentation with the application of CRFs (Conditional Random Fields) on the base of initial ELM training result.For conditional random field image segmentation technology is the research hotspot in the field of computer vision in recent years, the core idea of this approach is to construct a Probabilistic Graphical Model (PGM) with observation data, give a potential function corresponding with the model, according to CRFs learning and inference method, get the maximum probability of making the semantic tags as the final mark. The novelty of this model is that it takes pixel gradient characteristics, location characteristics, the front frame of the relationship between car video frame and the edge information into consideration, and takes it as the largest clique potential function to establish model. With the application of Broyden Fletcher Goldfarb Shanno (BFGS) and Loopy Belief Propagation (LBP), the model has been trained and inferred respectively. Proposed algorithm has been tested in public available data sets like MSRC (Microsoft Research Cambridge), ICCV09DATA and real scenes video, the recognition rate was 84.29%.The purpose of this thesis is to identify the associated region and region, as well as scenes semantic image segmentation in context information.