| License plate detection technology is an important part of intelligent transportation system, commonly used in traffic lawenforcement, parking management, automatic toll booths and other occasions. The license plate detection algorithm is researched in this paper, several methods are proposed to improve the performance of classifier and an embedded license plate detection system based on DM365 is designed. Details are as follows:Firstly, a method to improve the classifier performance by extending features is proposed. Because the extended features can enrich the original features, the license plate detection algorithm is designed by extended Haar features and Adaboost algorithm. The quick calculation method for extended Haar features is introduced and the training process of classifier is described. The detection frame is composed by multi-scale detection and windows merging. At last, a comparative experiment of basic Haar features and extended Haar features is conducted to evaluate the influence of extended features to classifier.Next, changing cascade structure can also improve the classifier performance. The Soft Cascade method is proposed to instead the traditional cascade method and a classifier based on Soft Cascade method is trained. Compared to the conventional cascade structure, Soft Cascade structure makes full use of the classification history information, resulting in better decision-making performance. At last, the classification performance of traditional cascade and Soft Cascade is evaluated.Finally, the completed license plate detection algorithm is transplanted into the embedded platform, several optimization method such as downsampling optimization, data type optimization, float-point operations optimization, software pipelining optimization, etc are carried out. After optimization, the time consuming of embedded license plate detection system is reduced greatly, while the system accuracy remains unchanged. The final system meets the needs of practical applications. |
PDF Full Text Request