Research On License Plate Recognition Technology For Surveillance Video

In recent years,the Intelligent Transportation System(ITS)has played an increasingly important role in construction of smart cities and urban traffic management.License plate information is an important indicator to distinguish different vehicles.Therefore,the license plate recognition technology plays an important role in the intelligent transportation system.Common license plate recognition technologies for smart parking lots and highways are difficult to promote to surveillance video scenes.On the one hand,general purpose target detection algorithms have poor performance in license plate detection at large angles.On the other hand,the license plate recognition algorithm based on single-character segmentation significantly reduces the performance of surveillance video scenes with uneven imaging quality and high noise.In view of the above challenges,the main work of this paper are as follows.Firstly,this paper improved and designed a key point-based license plate detection and location algorithm.The algorithm uses license plate vertex detection instead of license plate circumscribed rectangle detection.Taking into account the actual situation of the license plate in China,when the convolution of the feature diagram returns to the vertex of the license plate,a 3*1 convolution kernel is used to replace the commonly used 3*3 convolution kernel.Experiments show that compared with the general target detection algorithm,the key pointbased vehicle license plate detection and location algorithm proposed in this paper has an accuracy improvement of 34.6% when the IOU threshold is 0.95.The algorithm can also handle license plate data with some tilt angles in the video,which makes the algorithm more robust.Secondly,the license plate recognition algorithm based on convolutional neural network and LSTM is improved and implemented.This algorithm uses convolutional neural network to extract features for license plate recognition.The extracted features are segmented and concatenated to construct license plate feature sequences.In order to make full use of the associated information of the feature sequences,a stacked bidirectional LSTM is used for identification.This not only reduces the step of character segmentation in the general license plate recognition algorithm,but also increases the recognition accuracy by about 4.1%.Thirdly,based on the license plate key point detection algorithm and license plate recognition algorithm in the above research,this paper designs and implements a license plate recognition SDK.The license plate SDK interface is simple and the recognition accuracy is high.