Algorithm Design And Application Research Of Multi-Pose Vehicle Recognition

The high speed development of economy, bring people material life level increased significantly, people demand for vehicles is increasing, but at the same time, traffic environmental problems, traffic management technology issues are also increasingly prominent, in many ways, the traditional traffic management methods cannot meet the needs of the changing situation. Under this background, the role of intelligent transportation system is particularly prominent. And vehicle recognition system is a core technology in the field of intelligent transportation research, as well as digital picture processing, CV and pattern recognition, and other areas of the cross the hot topic of research, obtained the rapid development in recent years.This paper focus on the multi-pose of vehicle recognition research. In order to solve the existing vehicle recognition system mostly fixed posture vehicle models can only identify the problem, Using SURF (speeded up robust features) and vehicle recognition algorithm of multiple points of view. Under the different point of view, the algorithm of feature matching stability of statistics, obtained with the increase of Angle Scale-invariant feature transform (SIFT) extraction algorithm to extract the perspective of the characteristics of stability, within the scope of the 360 °, every 45° select a sampling point, eight Angle samples are selected; With the improved scale invariant feature extraction method, models for SURF some characteristic; Using Euclidean distance space and nearest neighbor search algorithm with the combination of feature matching method, and the nearest neighbor search algorithm to optimize design of k-d tree search process; Using PROSAC (progressive sample concensus) to eliminate false match. Through visual clustering and modeling method of multiple points of view, training the full range of vehicle models feature library, and in the form of an XML file stored offline.System testing process is mainly divided into training module testing and vehicle recognition test. Training module mainly models feature library building process, testing is mainly focused on the different features to eliminate false matching and purification methods on the influence of the feature library and inferiority. Vehicle recognition tests, with test images and models extracted feature set respectively with trained the feature library for all models feature matching, output highest matching models. And in a large number of tests under different conditions, analysis of conditions on the impact of the algorithm.The experimental results show that the proposed multiple posture model identification algorithm is feasible and effective, the algorithm is accurate recognition rate at 90%, has a good real-time performance and stability. To the design of similar algorithm has good reference value, for vehicle recognition in intelligent transportation provides new ideas in the field of application.