Research On The Algorithms For Vehicle Type Recognition

Vehicle type recognition is an important subject in the study of Vehicle Control System,and this subject is one of the main key technologies in anew product research and development of enterprise cooperation project In the thesis,we take car face area as the object to be recognized and conduct study on algorithms of recognizing vehicle type.Our work is summarized as follows:? Segmentation of car face area.The shape features and structure features of car face area can represent different type vehicles.According to the detected license plate location information,we cut the car face area.The car face area is the object of study in the following sections.? Extraction and recognition of car face features.For the car face area,wE propose two technical lines for car face features extraction and vehicle type recognition One is that we extract structural features of the car face and use SVM to classify different vehicles.The other is that we use deep learning method to extract features and classify different vehicles.On one hand,we study three kinds of structural features:the HOG feature,the SIFT feature and the feature of partial car face matching template.HOG feature represents gradient edge information of car face well,and SIFT feature shows good adaptability to scaling car faces,while the feature extracted by matching partial car face and the pre-defined templates can better express car lights information and so on.Then,we use the trained SVM models to recognize vehicle type.On the other hand,we use deep learning method to recognize vehicle type.We use single-layer network and multi-layer network to extract car face features and train the classifier models.? Decision fusion of vehicle type recognition.We obtain different kinds of feature expression of car face with different feature extraction methods,and complete vehicle type recognition through classifier models trained by SVM and deep learning method.On this basis,we conduct decision fusion of multi-models to further improve the vehicle type recognition accuracy.Firstly,we get the fusion result by combining the decision results based on HOG feature?SIFT feature and the feature of partial car face matching template together.Secondly,we put together the fusion result with the decision results of deep learning method Eventually,?we promote the accuracy.In the thesis,we implement the algorithms stated above by C++ language and Caffe framework,and carryout the prototype of vehicle type recognition system.Besides,we have also conducted detailed experiments to prove effectiveness and practicality of algorithms mentioned above.All the car face samples which are total of about 7800 and 16 categories are collectedfromthe real road traffic.Experimental results showthat the decision fusion accuracy of vehicle type recognition based on structural features is about 98.32%,and the decision fusion accuracy of vehicle type recognition based on structural features and deep learning method is about 99.20%.