Design And Implementation Of Traffic Video Surveillance Algorithm Based On Embedded Platform

Road transport has gone beyond the rail transport, became the most important ground transportation. It plays an very important role in the national economy and social development. With the continuous growth of China’s automobile production, accompanied by increasingly severe road congestion. Traditional transport system has difficulty to adapt to the current development of society. Intelligent transportation systems came into being to solve this problem. It can improve the reliability and safety of the transportation system, reduce energy consumption and pollution to the nature. In recent years, with the development of computer science, video based traffic information extraction system has become a popular research direction in this field.This paper studies a kind of traffic sullivance system based on embedded platform and practice a software for the system. The software includes two functions:the traffic statistics and vehicle license plate recognition. Software analysis and process video from cameras to identify the vehicle and the license. After that, it saves license plate images into device. For the saved images, the system will identify the license plate number of each. After processing, the data will be uploaded to the back-end server through the network.In order to realize the vehicle counting and license plate location, we chose the famous Viola-Jones object detection framework. The method is better than the method based on moving object segmentation when the vehicles ran slowly and even they stilling. Its disadvantage is the need for training offline, and differences of target within class should not be too large. License plate recognition based on convolution neural network, this method is popular in deep learning recently. Compared with the traditional template matching or structure based recognition, it is amore accurate and more robust methods. Its disadvantage is computationally intensive and also need to be trained offline.To training Viola-Jones object detection framework and convolution neural network, this paper uses a background modeling based moving object segmentation method to collect samples of vehicles, to expand the samples of characters uses elastic deformation. The corresponding part of the thesis described those two methods.The experimental results are listed in the experimental section. Sence our system is base on embedded, in addition, some adjustments of software for the embedded platform are described. Experimental results show that the method used here have a high license plate recognition rate and could do real-time traffic statistics.