Research On Vehicle Feature Recognition Algorithm Based On Convolutional Neural Network

With the rapid development of economy,vehicles are playing an increasingly important role in daily life.While bringing people portable and comfortable,they also bring great pressure to road traffic.In order to solve the increasing road traffic safety problems,intelligent transportation system came into being.As an important part of intelligent transportation system,vehicle type recognition system has been widely used in urban traffic flow monitoring,highway automatic toll collection,inspection and criminal investigation.It is of great significance to assist the existing vehicle recognition system and strengthen vehicle management.This thesis combines the Faster R-CNN network of convolution neural network to detect and recognize the vehicle image captured by road surveillance.Based on the original Faster R-CNN network model,an improved Faster R-CNN model is proposed for vehicle recognition.The main work is as follows.(1)Aiming at the problems of small target in vehicle image,single feature extraction in original network,insufficient feature extraction and low recognition accuracy of small target,a multi-layer feature fusion network structure is proposed,which adds deconvolution layer,normalization layer and concatenation layer to realize feature fusion of low-level features and high-level semantics,and improves the expression ability of features.The experimental results show that the results show that the optimized model can recognize the target better.(2)In view of the uneven distribution of positive and negative samples in vehicle images and the disparity in proportion,an improved online hard sample selection method is proposed.Traditional convolutional neural networks have low recognition accuracy when dealing with the problem of disparity in the proportion of training samples.In this paper,the focus loss function is used to select the hard samples,and the feature re-learning of the hard samples is strengthened,and improves the vehicle recognition accuracy of the model.(3)Aiming at the problem of high overlap of candidate box,violent deletion of candidate region and missed detection of target at fixed threshold,a soft non-maximum suppression method is proposed.When the merging ratio of the candidate box and the truth box is larger than the threshold set,by reducing the confidence value of the candidate box and continuing to compare,the candidate box can be better retained,and the recall rate and recognition rate of the model can be improved.In order to validate the effectiveness of the improved algorithm,a comparative experiment is conducted on BIT-Vehicle data set by data enhancement technology.The experimental results show that the improved method proposed in this thesis can improve the detection effect of the model without increasing the recognition time.Compared with the traditional vehicle recognition method,the improved algorithm proposed in this paper has the advantages of short recognition time,high recognition accuracy and low false alarm rate.It can locate vehicle target more accurately.