A Research On The Method Of Vehicle Detail Feature Recognition Based On Video

The detection and recognition of vehicle detail features is an important part of the Intelligent Transportation System (ITS). Among the many features, license plate is the most critical feature of vehicle, and its detection and recognition is of great practical value. Although there have been many mature license plate recognition system available, most of the products can only get good recognition results in certain scenes and has low license plate locating rate and recognition rate in complex environment, such as large tilt angle, uneven illumination, rain, snow, fog and so on. This paper focuses on solving the key and difficult points in the license plate recognition under the environment of large tilt angle. The concrete research contents are as follows:(1) In the stage of license plate location, the two level locating method is employed. In the stage of rough location, vertical edge features combined with color features are applied to extract candidate license plate region. The effect of Hough transformation and affine transformation on tilting and skew images is compared, and finally affine transformation is used to correct tilting or skew license plate image. In the stage of fine location, the projection features as well as Support Vector Machine (SVM) two classification algorithm are used to classify the candidate license plate images with two times in order to get real license images precisely. The experimental results show that with the proposed method the locating rate is 0.9074 in the environment of large tilt angle,0.9462 in the entrance bayonet scenarios,0.9500 in the freeway scene and 0.9595 in randomly shot images.(2) In the stage of license plate character segmentation, the method based on contours and Chinese character reconstruction is adopted, and the Chinese character in the license plate image is completely segmented as well as other characters. The experiment shows that the segmentation accuracy is above 98%with the proposed method.(3) In the stage of license plate recognition, the effect of Back Projection (BP) neural network and SVM algorithm in varied kinds of features is compared, and finally Histogram of oriented gradients (Hog) features and SVM multi-class classification algorithm are applied. And we get good recognition results. The experiment shows that with the proposed method the recognition rate is 0.9184 in the environment of large tilt angle,0.9765 in the entrance bayonet scenarios,0.9474 in the freeway scene and 0.8795 in randomly shot images.(4) Finally, a license plate recognition software system is implemented in this paper by using opencv and VS2012. The software system adopts multi-thread processing technology, and can handle both images and video files in an active state.The proposed method in this paper can effectively solve the problem of low locating rate and recognition rate in large tilt angle, and can also reach high rates in the entrance bayonet scenarios and freeway scene. Therefore, the propose method in this paper has certain practical value.