Research On Vehicle Detection Algorithm Based On R-CNN

Vehicle detection has important applications in assisted driving,traffic management,remote sensing images,etc.With the increasing number of vehicles in recent years and the increasing demand for vehicle detection technology,vehicle detection has become an important research topic in the field of object detection.The traditional object detection method can improve the accuracy of vehicle detection,but it cannot achieve good results in complex scenes,and there is a bottleneck in vehicle detection.The deep learning-based object detection method uses a large amount of image data to train complex network models,which can extract deeper and more effective object features in the image.The detection effect in complex scenes is better than the traditional method,and the performance of vehicle detection is improved.The object detection method based on deep learning is an important research topic of researchers,and has achieved very good research results,and has great potential in future research and application.Based on the research of vehicle detection and deep learning,this paper develops different algorithms to detect vehicles according to different sample sizes in various scenarios,which provides a new idea for the improvement of vehicle detection performance.The main research results are as follows:(1)For the detection time and storage space problem in the deep learning vehicle detection method,we replace the VGG network in the original Faster R-CNN algorithm with a MobileNet feature extraction network whose basic components are Depthwise Separable Convolution(DSC).Reduce the amount of parameters and calculations generated in network convolution by changing the process of convolution.In the RPN,the width and height of the anchor boxes obtained by the clustering Ground Truth(GT)is introduced,to get more accurate candidate regions faster,which is used as the basis for the target classification and the bounding box regression.The algorithm reduces the storage space of the network model to 42.52MB,and the time of vehicle detection of real road images and optical remote sensing images is reduced to 0.07s and 0.09s,respectively.(2)For the detection of small vehicles in optical remote sensing and real road images,we unified the multi-scale feature map extracted by Faster R-CNN algorithm feature extraction network MobileNet through the maximum pooling and deconvolution operations,and fuse the feature maps.The fused feature maps is used as shared feature map to optimize the situation of false detection and missed detection of small target vehicles in the object detection algorithm.Based on this situation,the detection performance of the small vehicles can be improved,and the average detection accuracy of the real road image and the optical remote sensing image vehicle detection is improved to 94.43%and 85.21%.(3)For vehicle detection in synthetic aperture radar(SAR)images,SAR image data is limited.Using deep learning methods cannot train an effective model,so that it cannot achieve good detection results.Based on the small sample,we propose a SAR image vehicle detection method based on feature fusion sparse representation model,which integrates vehicle detection into the Sparse Representation(SR)fusion framework.First,a set of residuals,for one specific feature,is first generated by performing the sparse reconstructions over dictionaries associated with the available set of possible targets.They are then normalized and further formed into a single residual sequence.After the collection of all residual sequences for all types of features,a linear fusion strategy is applied to the sequences to infer an optimal target estimate.The test results based on real scene data show that the proposed method achieves an accuracy of 97.17%for vehicle detection in SAR images.