Research On Vehicle Detection And Vehicle Recognition Based On Deep Learning

Vehicle detection and vehicle type classification are important research contents in the field of intelligent transportation.And with the development of artificial intelligence technology,applying deep learning to solve vehicle detection and vehicle type classification problems in the field of intelligent transportation has become a research trend.This paper studies vehicle detection and vehicle type classification based on deep learning methods.The main research contents are as follows:(1)Aiming at the problem that vehicle detection databases nowadays are few,this paper chooses the traffic environment around Nanjing University of Information Science and Technology as the research background,and then introduces a vehicle detection dataset in detail from the aspects of sample shooting and collection,selection of positive,negative and fuzzy samples,labeling and so on.The dataset divides vehicles into four types: car,truck,van and bus.It includes different scenes such as day,night,sunny,cloudy and rainy days.It also includes different vehicle orientations to ensure the diversity of vehicle detection samples.In addition,the Stanford Car-196 dataset has few pictures of each model and the dataset also exists the problem of sample imbalances.The study adds Gaussian noise,mirror processing and other data enhancement algorithms to expand the existing Stanford dataset.The above two datasets prepare for subsequent vehicle detection and vehicle type classification research.(2)Aiming at the situation that the original faster-rcnn manually set the parameters of anchors which cannot be well adapted to specific scenarios and the small target detection effect is also not good,this paper proposes a small target vehicle detection network based on improved faster-rcnn.By using the k-means clustering method,the positions of target frames in samples of the training set are clustered and replace the set 9 candidate windows with k cluster centers of good effects to improve the detection accuracy.In addition,the ROI Align algorithm based on bilinear interpolation is used to pool the feature map output by the feature extraction network and the position information output by the RPN network.This effectively improves poor detection of small targets due to pixel offset caused by the rounding operation of ROI pooling.Experiments show that the algorithm has a good effect in the test scenario of this article with high accuracy and robustness.It also has a certain practical value.(3)Aiming at the problem that the accuracy of current vehicle type classification algorithms are not high enough due to factors such as shooting angles and etc.This paper proposes a finegrained vehicle type classification method based on feature optimization and joint learning strategy.This article first uses the idea of transfer learning to select three commonly used CNN models and compare them after retraining on the training set of this article,and improve the model with the best performance.The feature optimization module SENet is added to improve the ability of extracting features of CNN.In addition,we use a joint learning strategy that combines softmax loss and center loss to improve the fine-grained vehicle type classification ability of the network.The experimental results show that the vehicle type classification method in this paper is superior to the existing algorithms and has the recognition ability under multiple angles.(4)In order to simulate the actual landing use of the algorithm,this paper uses tkinter which is a GUI design module of python to develop the vehicle detection interface and vehicle type classification interface.The two system interfaces call the above algorithm in the background to test the input samples and visualize the test results.The two interfaces avoid complex operations of commands and improve human-computer interaction and convenient operability.