Recognition Of License Plates Based On Video Sequences

With the development of China’s economy, the number of cars is increasing,traffic problem is becoming more and more serious, so the intelligent transportationhas become a research hotspot in recent years. License plates recognition is the keyand core technology in intelligent transportation system. It has a broad applicationprospect, such as road monitoring system, vehicle violation, accident investigation,automatic parking management, highway bridges and highway toll system, city vehiclemanagement and so on.The video vehicle identification system consists of four modules: moving vehicledetection, license plate location, license plate character segmentation and characterrecognition. Based on the deep study and research of the existing theory and algorithm,I effectively and perfected improved the key technology in the license platerecognition. The main contents of this paper are as follows：First of all, backgroundsubtraction techniques is used to detect moving vehicles, then Gaussian mixture modelis used to model the background of video surveillance scene.The advantage of thisprocedure is that it not only detect moving vehicles, but also remove the region beyondthe vehicle，and make the license plate positioning more accurate. A new algorithm oflicense plate location based on the combination of edge detection and prior knowledgeproposed。Firstly，the image was smoothed by operated gray stretch to reduce thenoise, secondly, the image was edged operator by edge detection, so we got thelocation of the license plate area location. Thirdly, after binarized of binary image, theimage was projected in the horizontal and vertical direction, we got the location of thelicense plate again. Finally, integration of the location of the two methods, we can getthe license plate from the image. In the license plate character segmentation module,an improved method is used to segment the character to reduce the effect of the licenseplate frame and noises. This method used fixed features of our license plate andvertical segmentation, we get characters by scanning boundary. This method is bettersegmentation of the characters. In the character recognition module, because the BP algorithm for training multilayered feed forward neural network has someshortcomings (eg. slow convergent rate, local convergence)，a learning algorithmbased on the classification is proposed to apply to the character recognition oflicense plate. The license plate characters are divided into three categories, characters,letters and numbers. According to the different characteristics, we used three sub-networks, and added the momentum factor when weight changing in training of theneural network for character recognition. Finally, we distinguish similar characterssecondary. The method improves the BP network learning speed and the accuracy ofcharacter recognition.Through the test of100monitoring car photographs on different road situationscaptured video, the results shows that the system can quickly, efficiently andaccurately identify the license plate, and achieve a higher license plate recognition rate.