Vehicle Target Detection Based On ResNet101

With the rapid development of China’s economy,the total number of vehicles has increased sharply,and the original traffic supervision system is relatively lagging behind.How to build a perfect,intelligent and accurate traffic supervision system suitable for the new situation has become an urgent problem to be solved.The traffic supervision system combines a variety of very important technologies,among which vehicle target detection is a critical one.Therefore,vehicle target detection has become a hot research topic in recent years,and has been widely concerned by scholars in various fields,such as artificial intelligence,computer vision,image processing and recognition.At present,there are mainly two vehicle target detection methods.The first method uses manual feature extraction combined with machine learning classification to detect vehicle targets,and the second one uses deep learning method to detect vehicle targets.Due to the disadvantages of the first method,such as low detection accuracy,slow target feature extraction,poor robustness and versatility,deep learning method,therefore,has been widely used in vehicle target detection in recent years.This method can automatically extract the features of image data,which are invariant to various transformations,deformations and complex environments,and has high detection accuracy and speed.However,this method is still defective when it is necessary to detect comparatively small vehicle targets,obscured vehicle targets and vehicle targets in rainy or snowy weather.In view of this,this paper uses Faster R-CNN(Faster Region Convolution Neural Network,Faster R-CNN)and Mask R-CNN(Mask Region Based Convolution Neural Network,Mask R-CNN)models of deep learning method to apply to vehicle target detection,and improves the models.At the same time,this paper studies the learning rate strategy of training model.After improving the method,the ability to detect comparatively small vehicle targets and vehicle targets in various complex scenarios has been greatly enhanced.The main tasks of this paper are as follows:1.MS COCO data set,Kitti data set and Pascal VOC 2007/2012 data set are analyzed,and on the basis of the comparative experiments,MS COCO data set is selected as the model training and testing data set of this paper.2.In order to improve the ability to detect vehicle targets,the network model based on regional proposal and the network model based on regression and classification are firstly studied,and Faster R-CNN model is chosen to be used as a vehicle target detection model in this paper.Then three kinds of feature extraction networks,i.e.Inception V2,Res Net50 and Res Net101,are studied,and Res Net101 is selected as the feature extraction network of Faster R-CNN.Finally,a large number of experiments are carried out on the learning rate strategy when training the model and the learning rate strategy with the highest experimental accuracy is found through the accuracy comparison.3.The anchor frame size of regional proposal network of Faster R-CNN is adjusted.Thus,the ability to detect comparatively small vehicle targets is enhanced and at the same time,it has certain detecting ability for vehicles in complex environments.4.Two improved models of Faster R-CNN are studied.The Mask R-CNN network model with Res Net101 as feature extraction network is decided to adopt for vehicle target detection.The anchor frame size of regional proposal network is adjusted to further enhance the detecting ability of comparatively small vehicle targets,and the robustness of the model is also strong.