| With the vigorous development of driverless cars,various countries are racing to participate in the technology research and development of driverless cars.At present,some countries have put unmanned vehicles on trial.Environmental perception is a vital module in the driverless car system,and it is also the basis for subsequent behavioral decision-making,path planning,and other modules.Object detection technology is a significant technology in the environment perception module.Identifying and locating objects in images is the purpose of object detection.Because the technology can quickly identify objects and accurately locate them,it is applied extensively in the field of unmanned driving.In the past few decades,because of the great potential of deep learning technology,object detection technology has achieved obvious effects.In recent years,R-FCN has been broadly used to deal with the issue of object detection,and the experimental results are relatively ideal.R-FCN is improved based on Faster R-CNN.R-FCN uses Position-Sensitive ROI pooling instead of ROI pooling to obtain the score of the region,which effectively improves the sensitivity of the network to object location information.Particularly,R-FCN uses a shared full convolutional network to obtain a shared convolutional feature map,which realizes computing sharing.Although the object detection method based on R-FCN has made great progress in accuracy and speed,there are still many challenges.However,the research of R-FCN did not consider the following two points.First,due to the unreasonable design of the original anchor boxes and the excessive number,the object recognition and positioning abilities are poor,and the calculation speed is slow;second,the softmax classifier has the poor fitting ability when distinguishing between positive and negative region proposal,which affects the region proposal generates the classification effect of the network.In response to the above two issues,this paper proposes a vehicle detection algorithm based on improved R-FCN and SVM,referred to as IR-FCNSVM.The main research content of this paper includes the following two parts:(1)Aiming at the issues existing in the original anchor boxes,novel anchor boxes are designed in this paper.6 sizes of anchor boxes based on the aspect ratio of the golden section coefficient are used to replace the original 9 sizes of anchor boxes to solve the problem of weak positioning ability in object detection.(2)In the Region Proposal Network,softmax,which is used to distinguish positive and negative region proposals,is replaced with Support Vector Machine.Whale Optimization Algorithm is used to optimize the parameters of SVM to improve the fitting ability of the classifier to distinguish between positive and negative region proposals and then improve the classification ability of the algorithm.This paper conducts experiments on the PASCAL VOC and KITTI datasets to verify the effectiveness of the algorithm in this paper.It can be seen from the experimental results that the algorithm has a better classification and recognition effect.In addition,the algorithm is compared with several other detection algorithms,and the performance of this algorithm is better.The value of the relevant parameters of the algorithm in this paper can be determined based on specific experiments,and the algorithm has good flexibility. |
