The Study On Visual Slam Of Mobile Robot

Robots have been widely used in many fields such as aerospace,disaster rescue,riot control,warehousing and logistics.How to calculate the position of robots through sensors and build environmental maps is a key technology for robot intelligence.SLAM(Simultaneous Localization and Mapping)technology has emerged to solve this problem.In recent years,with the development of computer vision and camera technology,using vision sensors to build more robust SLAM systems has gradually become a research hotspot.As the vison front of SLAM,the visual odometry,which processes the camera's raw data,plays an important role in the initial positioning of the robot.Loop closure detection also plays an irreplaceable role in reducing the accumulative error of the robot.So this paper takes the visual front and loop closure detection in the visual SLAM algorithm as the research object,and explores the method of using neural network to learn image features for loop closure detection.The main research content is as follows:Firstly,research on feature points evenly distributed.Aiming at the problem that the traditional image features are unevenly distributed in the image,the improved quadtree is used to manage the rough extracted feature points,so that the feature points can be evenly distributed in the image.By comparing with the current mainstream feature extraction algorithm,the effectiveness of the algorithm for improving uniformity is proved.Then the effectiveness of the algorithm to improve the positioning accuracy is verified by trajectory accuracy experiments.Secondly,research on image representation based on neural network.In view of the problem that the traditional word bag model does not have a good performance on the image representation between different images.On the one hand,the convolution autoencoder is designed for image representation.And the autoencoder is empowered with the ability to remove image noise and extract sparse features by adding noise constraints,sparse constraints,local connection layers and the appropriate loss function.On the other hand,the pre-trained neural networks are used to represent image.Then the similarity score of different images were calculated with BoW and the algorithms designed in this paper.The superiority of the algorithm in image representation is verified by image similarity comparison experiment.Finally,loop closure detection based on convolutional neural network.The public data set are used to verify the performance of the different loop closure detection methods include convolution auto encoder,pre-trained models and BoW.Then the PR curve is obtained by calculating the precision and recall under different threshold to verify the validity of the proposed method for loop closure detection.In summary,this paper mainly studies the feature extraction and loop closure detection in the visual SLAM algorithm,designs an ORB feature uniform extraction strategy and a loop closure detection method based on convolutional neural networks.And then the experimental verification and analysis of the above algorithm is carried.The experiments prove that the improved algorithm is effective for improving the uniformity of ORB features.At the same time,the neural network is used to represent the image and improve the loop closure detection performance,which has certain theoretical significance and engineering reference value to improve the positioning accuracy of visual SLAM algorithm.