Research On Vehicle Classification And Recognition Method Based On Image Convolutional Neural Network And Deep Learning

Vehicle type recognition using visible image is of great research and practical value in application scenarios such as lane monitoring,fake plate vehicles detection,highway ETC charges,autonomous driving,military vehicles precision strikes,etc.As an important research direction,vehicle type recognition methods based on deep convolutional neural network have better performance than other machine learning methods.However,factors such as multi-scale,small dataset and poor embedded real-time performance affect the application effect of traditional deep convolutional neural network in vehicle classification.In view of these problems,the vehicle type recognition method with embedded real-time used in multi-scale small dataset was studied based on deep convolutional neural network.Firstly,according to the characteristics of vehicle type recognition,the overall scheme of offline training and learning + online recognition in embedded platform is designed.The fast image preprocessing algorithm,feature extraction method,classifier and parameter updating method in vehicle classification system are analyzed,designed and selected.In view of different vehicle angles,the scheme of small-angle random rotation and random cutting is selected and designed.In view of illumination difference,a numerical normalization algorithm is introduced to normalize the image brightness by artificial mean and standard deviation.In view of the scale variation of the same vehicle feature in images of different scales,a multi-scale branch convolution feature extraction algorithm is proposed,which uses different branches to extract the initial convolution features of different sizes.An adaptive learning rate attenuation method based on the product of loss function and iterations is designed.Secondly,the hardware platform of vehicle type recognition system is built.The training and testing software for vehicle classification algorithm on offline training platform,simplified model on embedded platform,graphical user interface,system circuit control software are designed and implemented.Finally,the multi-scale,multi-angle small dataset is constructed.The performance test and validation experiments of the vehicle classification method designed in this paper are taken in the dataset.On the training platform,the accuracy of 200 test datasets is 90%.On the embedded platform,the transplanted simplified model achieves 86.5% accuracy in 200 test datasets and23 pieces per second classification speed,meeting the real-time and accuracy requirements.