| At present,traffic video surveillance plays an important role in transport efficiency,traffic safety and emergency response,but the actual environment is complex,such as shadow,moving target occlusion,channel water wave,etc.The traditional background modeling algorithm is limited to its inter-frame pixel change Mechanism,although simple and real-time,but often difficult to solve the above interference factors.In addition,license plate recognition also has a similar situation.To this end,the paper combines the "ship network" national special project and Jiangsu traffic internet of things demonstration project and other projects.The research shows that it has important theoretical and practical value.This paper describes the current situation of difficulties or deficiencies in traffic surveillance video,including the interference factors such as shadow,occlusion and license plate recognition,etc;and expatiates the common video detection mechanism such as background modeling,optical flow and so on.The problem of vehicle video detection is discussed.For example,GMM modeling is used to extract the problems such as hole and noise,shadow recognition caused by false detection or missed detection,background modeling,And then discuss the initial concept of solving difficulties or deficiencies.In this paper,a series of algorithms are presented for the traditional video detection methods,including:In order to solve the problem that the traditional shadow detection algorithm is difficult to distinguish the boundary of the moving target that cause the similarity between the color and the texture information,the paper analyzes the HSV color feature and the LBP texture invariance extraction shadow principle,and explains and improves the SE-CT Shadow removal algorithm,a new method named GA-HT algorithm is proposed,which is based on color and texture feature and gradient filling.The shadow detection algorithms are analyzed and compared,and the validity and feasibility of GA-HT shadow removal algorithm are verified.Aiming at the problem that the DPM can not deal with the target partial occlusion problem caused by the detection error of the parts detector,the paper proposes a target occlusion processing algorithm based on DPM-CRF combination.The algorithm analyzes the principle of DPM model,extracts the target context information and the spatial configuration information between components,through the DPM model to train and optimize the CRF parameters,and then achieves the final target detection by the confidence propagation algorithm.Experiments show that the algorithm can solve the problem of vehicle occlusion.Aiming at the problem that the efficiency of the traditional license plate recognition algorithm is low in the complex environment and the cost of the calculation is high and it is difficult to distinguish the similar characters effectively.The paper analyzes the KNN algorithm and the SVM classification algorithm,and proposes a license plate recognition with KNN and SVM model algorithm.The algorithm uses KNN as the first classification model to identify the normal and regular characters,and then uses the RBF kernel SVM classification model to deal with the similar character classification problem.Finally,the experimental verification improves the accuracy of the license plate recognition.In summary,the innovation and characteristics of the paper are:●It proposes GA-HT shadow removal algorithm to remove the moving target(vehicle,ship)shadow,improved the accuracy of the target detection.● It optimizes the discriminant component model,and puts forward the DPM-CRF combination model,which solved the problem of partial occlusion of moving target,inhibited the fusion of adjacent target detection,and improved the accuracy of target detection and traffic flow parameters.● It improved a single SVM license plate recognition model,joined the KNN classification algorithm,to solve the problem of similar character false positives,improve the efficiency and accuracy of license plate recognition. |
PDF Full Text Request