Research On Visual SLAM Method Based On Feature Points Of Convolutional Neural Network

The camera moves in an unknown environment and autonomously determines its own position and spatial map construction.It is the main problem to be solved by visual real-time localization and mapping(SLAM).At the front end of visual SLAM,feature point method and direct method are the current mainstream methods,but these two methods have drawbacks.The feature point method SLAM has the problems of low matching efficiency,poor matching effect,and great influence on image quality.The development of deep learning in the field of vision is a must,but visual SLAM has a lot of algebraic optimization problems.It is difficult to design an end-to-end SLAM model with excellent results using deep learning methods.This paper avoids end-to-end SLAM and uses deep learning methods to solve the problem of feature point detection in visual SLAM,which is also the advantage of deep learning methods.This paper trains a feature detector based on a convolutional neural network,and further uses this feature detector to implement a monocular visual SLAM system,and adds related algorithm optimization.The main research contents are as follows:(1)For traditional hand-designed feature point extraction and matching,which are easily affected by factors such as illumination,viewing angle,noise,and image blur,this paper uses deep learning to solve the feature point detection problem.The traditional keypoint features are fused,and the robust optimization is targeted;the VGG-style fully convolutional neural network improved by MobileNet lightweight techniques is used as the encoder,and two parallel decoders are used to detect keypoints and Calculate descriptors,use dual network architecture to learn image pairs,and use cross entropy loss(Cross Entropy Loss)and hinge loss(Hinge Loss)to train key points and descriptors respectively.The experimental results prove that the MVP feature detector has a certain performance improvement compared with the traditional method,which lays a good foundation for the subsequent application of deep learning feature detector to SLAM.(2)Using the trained MVP feature detector,from the front-end visual odometry(Visaul Odemetry,VO)and closed-loop detection two modules,the use of MVP features in the SLAM system has been optimized and improved by algorithms.At the front end,a VO based on the MVP feature detector is constructed,and the quadtree algorithm is used to manage the key points,and the non-maximum suppression algorithm is used to filter the key points with the highest confidence of the child nodes,and the floating point descriptor is used Binary processing,and use an improved two-way nearest neighbor matching algorithm to match feature points;in the closed-loop detection module,Kmeans clustering algorithm is used to train a dictionary based on MVP features,and ktree representation of the dictionary is used to reduce the complexity of the word search algorithm.degree.Connect the VO and closed-loop detection modules to the back-end optimization and mapping modules to form a complete monocular SLAM system.The experimental results show that the optimization strategy for key points and descriptors improves the operating efficiency of the system;the MVP feature detector trained by the deep learning method exhibits higher accuracy and better performance in the VO module and SLAM than the ORB feature detector Real-time.