Research Based On Deep Learning For Autonomous Driving Monocular Vision Object Detection Technology

The convenience of people's daily life has been improved by increasing the number of cars.However,a lot of attention is paid by people in traffic jam,environmental pollution and traffic accident.In recent years,the electronic information and computer technology have developed rapidly.With the application of artificial intelligence technology in vehicle,autonomous driving vehicle will become one of the effective ways to reduce the number of traffic accidents.Environmental perception,precise positioning,path planning and control-by-wire are the main technologies for autonomous driving vehicle.The foundation of these key technologies is environmental perception which is mainly responsible for the collection of environmental information and object detection.Monocular camera is widely used because of its simple structure and low computational complexity.Because traditional monocular vision environmental perception methods mainly rely on prior knowledge to design models,thus the generalization ability of these methods is not strong.However,the model generated by the monocular visual environmental perception methods based on deep learning can be applied to many kinds of objects.Therefore the generalization ability of these methods is promoted.Consequently the technology of autonomous driving monocular vision object detection based on deep learning is the research emphasis in this paper.Recently,excellent experimental results are achieved by Faster Region-based Convolutional Network(Faster R-CNN)in solving the problem of object detection.The region proposals can be effectively generated by Regional Proposal Network(RPN)of Faster R-CNN with different scales and aspect ratios.Three monocular vision object detection methods are proposed based on Faster R-CNN algorithm in this paper.The main works and improvements of this paper are presented as follows:(1)An object detection method based on Multi-strategy Region Proposal Network(MSRPN)is proposed.A novel skip-layer connection network is designed for combining multi-level features and boosting the ability of pooling layers.Thereupon,the quality of region proposals is strengthened.Secondly,improved anchor boxes are introduced with adaptive aspect ratio and evenly distributed scales.In this way,the number of predicted region proposals for detection is seriously reduced and the efficiency of object localization is increased.Particularly,the capability of small object detection is enhanced by applying the first and second improvements.Thirdly,classification layer and regression layer are unified as a single convolutional layer.Furthermore,the model complexity of output layer is reduced.Thus,the speed of training and testing is accelerated.Fourthly,the bounding box regression part of multi-task loss function in RPN is improved.Consequently,the performance of bounding box regression is promoted.(2)Based on part of MSRPN improvements,an novel object detection method based on Enhanced Region Proposal Network(ERPN)is proposed.The improvements of ERPN are presented as follows.Firstly,our proposed deconvolutional feature pyramid network is introduced to improve the quality of region proposals.Secondly,novel anchor boxes are designed with interspersed scales and adaptive aspect ratios.Thereafter,the capability of object localization is increased.Thirdly,a Particle Swarm Optimization(PSO)based Support Vector Machine(SVM),termed PSO-SVM,is developed to distinguish the positive and negative region proposals.Fourthly,the classification part of multi-task loss function in RPN is improved.Consequently,the effect of classification loss is strengthened.(3)Part of the parameters in RPN is assigned by prior knowledge.Therefore underfitting problem is likely to appear on the training model of RPN.In other words,the generalization ability of RPN is not enough.Increasing parameters is an effective solution for this problem.Thereupon a Strengthened Region Proposal Network(SRPN)is designed to expand the exploring space of RPN.Acquiring the optimal parameter values of SRPN cannot be solved in polynomial time.In order to solve this problem,a PSO and Bacterial Foraging Optimization(BFO)based learning strategy is introduced to optimize the parameters of SRPN.In order to verify the effectiveness of the object detection methods proposed in this paper,the general object detection data set and the autonomous driving object detection data set are applied to train and test these methods.Our proposed algorithms can achieve good object detection results quickly and accurately.In other words,our proposed methods can be effectively applied to the general object detection and the field of autonomous driving.Therefore,autonomous driving monocular vision object detection technology based on deep learning proposed in this paper has important theoretical significance and application value.