Research And Implementation Of Vehicle Multi-feature Recognition Method

Along with the in-depth development of our country urbanization,the urban road traffic is faced with enormous pressure and challenges,all sorts of problems emerge in endlessly,such as serious road congestion and frequent occurrence of traffic accidents,these problems had a certain influence on social and economic development,also for city public security management brought a lot of trouble.In order to alleviate the pressure of traffic administration,and to provide a more convenient travel conditions,intelligent transportation system arises at the historic moment.In this paper,we researched the vehicle multi-feature recognition that is an important part of intelligent transportation system,and it mainly includes the recognition of vehicle number plate,model and color.This paper focuses on the research and implementation of vehicle multi-feature recognition algorithm,launched the following work:(1)We researched a vehicle detecting method based on aggregation channel feature,this feature contains three LUV color space channels,one gradient amplitude channel and six histogram of oriented gradien channels,which not only contains the whole target contour feature,but also contains the target local gradient feature.This feature shows a good accuracy in single view image such as highway monitoring data and car driving recorder monitoring data.For the low contrast image of vehicle,we researched an image enhancement algorithm,which can effectively alleviate the impact of environment and lighting,improving the recognition accuracy of vehicle number plate,model and color indirectly.(2)We researched a vehicle number plate locating method based on color and character feature,which fully combines the advantages of color and character.In the test on self-built database of the complex scene,the recall rate is 93.12% and the accuracy rate is 94.49%.We eliminate the false license plate by using the the SVM classifier that trained with the LBP characteristics of the number plate area.For the character recognition,we design a multi-level SVM strategy to reduce the false recognition rate of the easily confused characters,and the recognition accuracy of single character is 96.15%.Experiment proved that the algorithm is applicable to a variety of complex scenes and has a good robustness.(3)We discussed a vehicle model recognition method based on deep learning,solved the problem that the traditional machine learning algorithms can't carry out a detailed classification of vehicle model.Adopting the tactics of fine-tuning the pre-trained network model,we get a test results in vehicle model database as follows: the top-1 error rate is 28.2% and the top-5 error rate is 12.9%.Then,we improved the method of vehicle color recognition based on vehicle number plate's location.We discussd a vehicle color recognition method based on deep learning,and this method can effectively avoid the classification errors which caused by the ROI location errors.We researched the influence of color space on vehicle color recognition under the background of deep learning.We conclude that RGB color space has the better classification accuracy in the vehicle color database,and the average accuracy rate reached 93.77%.