Vehicle Recognition Based On Multiple Features Of The Image

Vehicle identification is an integral part of modern traffic control systems.It is also the basis for other intelligent tasks such as traffic management,driverless driving,tracking suspects,and behavior analysis.In the case of highway complex factors,there is a problem that vehicle type misjudgment and vehicle color change are not easy to identify.Traditionally,the main identification method for “gray license” vehicles is to extract the underlying features such as the texture,color,and size of the vehicle,and then use various classifiers to identify the vehicle.The main recognition methods are image matching,three-dimensional model and common machine learning methods.These methods are basically suitable for small sample data,but they are weak for the current big data.Therefore,this paper uses the convolutional neural network method based on image multi-features to identify the vehicle,which avoids the problems of traditional vehicle identification methods such as large manual work volume,weak resistance to natural factors,and low recognition efficiency.Overcome the impact of various differences such as different vehicle scenes,and at the same time overcome the difficulty of acquiring the underlying features.The main research work is as follows:1.The establishment of image data sets for expressway HD vehicles(model color logos).High-definition vehicle image data collection is based on images captured by high-definition devices in the real monitoring of each bayonet of the expressway,and the established model standard model library pictures are based on the principle of universality,extensiveness,and convenience for experimental extraction and implementation.,Select a training set and verification set according to a certain percentage.This data set serves as the basis for subsequent vehicle identification studies.2,sample data set preprocessing.To ensure the authenticity of the data,manually label the data set and then use the tools of categorizing data in NVIDIA digits to normalize the high-definition color image of highway vehicles into a uniform size to create a deep learning framework that can be quickly Reading format,according to a certain percentage of training set and verification set.Due to the limited raw data,experimental data sets need to be classified and specific,and some irregular images cannot provide effective vehicle information content.Moreover,the problem of vehicle identification has a high requirement for the generalization ability of the classification model.Therefore,the original data is expanded 3 times,and different experimental results are obtained based on the difference in the amount of experimental data.3.The traditional shallow learning method and convolutional neural network method are used for vehicle identification for small sample data volume and a large number of vehicle image data sets,and the experimental results are compared and analyzed.The experimental results show that the average recognition rate of the three models under the Caffe framework is higher than other methods of vehicle identification,and has obvious experimental results.