Research And Implementation Of SLAM Based On Deep Learning

With the development of computer vision technology,vision-based closed-loop detection technology is gradually applied to simultaneous localization and mapping(SLAM)of robots.Bag-of-words model is the most commonly used closed-loop detection method at present.Because the model needs artificial design features,it has great application limitations in the detection process.The extraction of these features takes a lot of time and is easily disturbed by external environmental factors,which will affect the accuracy of closed-loop detection.Deep learning technology is to learn image data features from classifiable raw data and classify them.In essence,closed-loop detection can be regarded as a classification problem,so deep learning technology provides a new method for closed-loop detection.Through the research and implementation of SLAM based on deep learning,this paper mainly accomplishes the following tasks:(1)The construction of software and hardware platform of mobile robot.The validity of the software and hardware development platform is verified by simulation experiments and practical experiments,which provides an experimental platform for the subsequent SLAM implementation based on deep learning.(2)Aiming at the problem of Closed-loop Detection in SLAM,the traditional closed-loop detection method of visual word bag is used to carry out experiments.TUM data sets are used.The experimental results show that the similarity score of the first and tenth images is about 5.25%,while that of the other images is about 2%.The error of map construction corresponding to the closed-loop detection model is large.(3)In order to improve the accuracy and real-time performance of Closed-loop,detection in mobile robot synchronous positioning and map creation,YOLO convolutional neural network is applied to closed-loop detection,and tested in YOLO 9000 model using laboratory data.The experimental results showed that under the same rate,the accuracy is high,and the speed of feature extraction is increased more than 140 times,which is higher than the traditional closed-loop detection based on visual word bag.(4)For a further improvement of the accuracy of closed-loop detection,the Closed-loop Detection Based on YOLO 9000 model is improved previously.By increasing the convolution layer and changing the number of convolutions,the Closed-loop Detection Experiment of the improved model is carried out again.Compared with YOLO 9000,therecall and accuracy are improved on PR curve,so the improved convolution neural network greatly improves the closed-loop detection and accuracy of measurement.The improved YOLO 9000 model is applied to SLAM to realize map repositioning and map building.By a comparison,the error in the map is obviously reduced after using the improved model.There are 50 figures,3 tables and 64 references contained in this paper.