Study On Several Key Technologies On Vehicle Recognition & Tracking Based On Image Analysis

Intelligent transportation system is a research hotspot in recent years, and digital image process and analysis technologies have been widely applied in this field. In this paper, some important technologies, such as, license plate recognition (LPR), moving vehicle detection, vehicle classification and vehicle tracking, are researched and discussed based on the practicle projects.For LPR, it includes three modules such as license plate orientation, character segmentation and character recognition. There are several license plate orientation algorithms proposed in this paper, for example, gray license plate image orientation, size adaptive license plate orientation and rapidly orientation based on template matching. Then, as for character segmentation, this paper adopted two algorithms, the one is based on connected region detection, and the other is based on projection. In addition, it is discussed primarily for license plate orientation and character segmentation in the evening. Then, about character recognition, this paper discussed the RBF neural network. The size adaptive license plate orientation method overcomes the problem how the morphologic structure element was adjusted adaptively to the size change of license plate. The experiment indicates that it is feasible to adopt this algorithm in LPR system to achieve accuracy and adaptabilityVehicle detection is often a challenge problem, especially for vehicles similar to the background. In this paper, a moving vehicle detection method is presented based on Kernel Density Estimation (KDE) model and edge model. The method can repair the imperfect region to some extent. Also, this paper presented how to compute the elements of KDE algorithm. The experiments results show that the vehicle detection algorithm can achieve accuracy for detecting moving vehicles.In additioin, in this paper, vehicle classes are recognized according to Stochastic Particle Swarm Optimization (SPSO). Vehicle features, such as vehicle size, shape information, contour information and edge information are extracted for SPSO training. Each vehicle class is encoded as a centroid with multidimensional feature and SPSO is employed to search the optimal position for each class centroid based on fitness function. Single classifier and decision tree classifier are utilized for classification. The algorithm's evaluations on video image series, moving vehicle detection, vehicle classification and tracking are respectively conducted. The results show that SPSO ensures a promising and stable performance in recognizing these vehicle classes, decision tree outperforms single classification, and the particle filter algorithm can achieve accuracy and real-time for tracking moving vehicles.Furthermore, this paper proposed a vehicle classification algorithm based on cloud model. Cloud model is a new theory which can express the relationship between randomness and fuzziness. Firstly, a method was adopted to detect the moving vehicles which are very alike to background. Each vehicle class is expressed through cloud model parameters, such as Ex (expectation), En (entropy), with multidimensional feature. And cloud classification model is employed to judge the optimal class for each vehicle. Furthermore, attribute similarity is introduced to judge the weight of each feature in classification. Single classifier and decision tree classifier are utilized for classification. The algorithm's evaluations on video image series, moving vehicle detection, vehicle classification are respectively conducted. The results show that cloud model ensures a promising and stable performance in recognizing these vehicle classes, and the algorithm can achieve accuracy and real-time.Moving vehicles tracking is also often a challenge problem, especially for vehicles similar to the background. In this paper, a moving vehicle tracking method is presented based on meanshift and particle filter algorithm. Then, an improved multi-feature fusion meathod using improved color feature model and edge feature model was presented for objects tracking under complex scene. The experiment indicates that it is feasible to adopt the above algorithm under the multi vecicles tracking scene to achieve accuracy.