3D Object Detection Based On Deep Learning

In recent years,the theory and research of vehicle autopilot have attracted wide attention in the field of intelligent traffic.Road object detection has been rapidly developed as one of the key links in the realization of autopilot.Traditional road object detection usually integrates visual positioning,pattern recognition,image processing and other related technologies to achieve 2D detection of road objects.Automatic vehicle driving not only requires more accurate and efficient detection of road objects,but also needs to obtain the depth information of the detected objects in order to reduce the risk factor of automatic vehicle driving.The traditional road object detection algorithm can't obtain the depth information of the object.Therefore,considering that the on-board camera is mostly a monocular camera,searching for a 3D object detection scheme based on a monocular image is an important research topic in the field of vehicle automatic driving.Traditional object detection algorithms based on the S-SVM(Structures Support Vector Machine)classifier have achieved very impressive results over the past dozen years.With the rapid development of large-scale training data and high-performance computers,deep convolutional neural networks with massive network model parameters can be efficiently fitted.We combines traditional object classification algorithms with deep learning and applies the advantages of S-SVM classifiers to deep convolutional neural networks,which has achieved very good detection results.Based on the S-SVM classifier,a preliminary detection of road objects is performed.First,multiple 3D templates are used instead of the popular SS(Selective Search)to generate 3D bounding boxes.At the same time,the number of region candidate frames is reduced by discarding a region candidate frame with a low prior probability.Secondly,we constructs the objective function by extracting the features of the 3D bounding boxes,and uses the S-SVM classifier to train the input feature weights.Finally,based on the S-SVM classifier,rough detection of the road object is completed.The detection results of the S-SVM classifier are migrated to deep learning by Fast R-CNN,which can greatly improve the detection accuracy of the 3D bounding boxes.First,we performs NMS(Non-maximum suppression)on the 3D bounding boxes from the S-SVM classifier,and then inputs them into the extended Fast R-CNN network structure for training.This network used by Fast R-CNN is a deeper VGG16 network.After the conv5 layer of VGG16 network,the network structure is divided into two branches,one to calculate the characteristics of the 3D bounding boxes,and the other to calculate the characteristics of the 3D bounding boxes that is expanded by 0.5 times.In this thesis,multi-task loss is used to predict the object category of the 3D bounding boxes,and the accuracy of object detection is improved by regressing the size and direction of the 3D bounding boxes.In order to verify the effectiveness of the algorithm used in this thesis,we first test the method of generating 3D bounding boxes based on 3D templates on the KITTI data set.By comparing the recall rates of various methods,we can prove that the method of generating 3D bounding boxes is superior to SS and other excellent algorithms.Then,the detection results obtained by using deep learning are compared with the world's leading road object detection algorithms.The experimental results show that this method can achieve the same or even higher recognition rate.It is proved that this algorithm has the potential for future vehicle autopilot applications.