| As the development of intelligent robot,local navigation and positioning technology is more and more useful,which does not rely on Global Navigation Statellite System.Simulataneous Localization and Mapping（SLAM）,which is the key technology of local navigation and positioning,is the strategic highland competing by major countries in the world.Using SLAM technology not only makes robot more intelligent,but also has an important influence on national defense.As a result,SLAM is a popular research topic in the field of robotics.In this paper,taking visual SLAM using monocular camera as research object,the front-end part of visual SLAM is researched,including Visual Odometry（VO）and loop-closure path detection.Research on using deep learning to improve the pose tracking accuracy of monocular VO,and improve the robustness of monocular VO under condition of different camera motions（including rotation-only,station-only,fast moving,variable speed）,and improve the detection precise of loop-closure detection for visual SLAM,and eliminate the estimated trajectory drift cased by uncertainty of monocular scale.The aim is to improve the accuracy of camera pose estimation and recover camera motion trajectory.The main research work of this paper is summarized as follows:1、As for the current end-to-end visual odometry which based on deep learning,such as CNN-VO and RCNN-VO,they have the problem of insufficient accuracy in estimating the camera pose.In this paper,a Deep Siamese Convolutional Neural Network（DSCNN）is designed,and an end-to-end visual odometry is implemented using DSCNN.As for CNN-VO and RCNN-VO,their network belongs to a linear network framework,which only considers the positive pose relationship between adjacent images,while for DSCNN-VO,it not only considers the positive pose relationship,but also takes the reverse pose relationship into account.In this way,it means that an additional constraint for artifical neural netwrok is added,so that the network training can converge to a better weight model.As a result,DSCNN-VO has better pose estimation accuracy and more robust for generalization ability.2、Although the accuracy of pose estimation for end-to-end VO can be improved by designing network architecture,and the generalization ability of network can also be improved.However,it still cannot fundamentally solve the problem of insufficient pose estimation.In order to solve this problem,designing a hybrid visual odometry which is based on deep learning and classical pose estimation methods,named DL-Hybrid VO.The VO uses two types of network structures:dense optical flow network and dense depth network,the former one is used to extract a dense optical flow field between adjacent frames,the latter one is used to extract a dense depth map from RGB image.A redundant pose estimation scheme is designed for DL-Hybrid VO,the purpose is to improve the robustness of DL-Hybrid VO,so that DL-Hybrid VO can still work stably in the condition of rotation-only or station-only motion,and fast motion of the camera;DL-Hybrid VO includes a monocular scale estimation module,which can estimate the accurate monocular scale online,in this way,the DL-Hybrid VO can greatly eliminating the trajectory drift phenomenon caused by the uncertainty of monocular scale.According to experimental analysis,the proposed DL-Hybrid VO algorithm can successfully recover camera’s motion trajectory when the camera’s motion speed is about 110km/h,while the traditional optical flow-based VO can only successfully recover camera’s motion trajectory when the camera’s motion speed is about 55km/h.3、The traditional visual SLAM algorithm uses the bag-of-words（BoW）method to detect loop-closure paths.However,as for BoW mehtod,it is not easy to train a perfect visual dictionary,because the number of clusters is a hyperparameter,if this parameter is setted not suitable,the features will not be clustered.The BoW method only considers the presence or absence of features in the image without considering the location,as a result,as for images with similar scene environments,the phenomenon of misjudgment loop-closure is prone to occur.In response to these problems,this paper proposes a loop-closure detection algorithm based on CNN network,named DL-LCD,which does not depend on a specific network,any CNN network trained for scene classification can be used.For DL-LCD,the effects of using different CNN networks,different CNN features,and different parameters values on the loop-closure detection results are analyzed through experiments,and it is concluded that the DL-LCD using the Inception4b CNN features of the GoogleNet can detect the largest number of loop-closure frames,while the DL-LCD using the Conv5 CNN feature of the AlexNet takes into account the operation time and the number of detected loop-closure frame,as a result,it is suitable for working in visual SLAM.The DL-LCD can work online and perform stably in different scene,and there is no need to modify the algorithm parameters according to the changes of the scene.Under the condition of 100%precision rate,the DL-LCD has a higher recall rate than traditional scene recognition algorithms.4、Combining the research results of this paper,a deep learning-based monocular SLAM algorithm is designed,named DL-SLAM.It is implemented by multi-threaded programming and adopts a multi-module design framework,which mainly includes three functional modules:visual odometry module,loop-closure path detection module,and back-end trajectory optimization module,visual odometry module uses the DL-Hybrid VO scheme,the loop-closure path detection module adopts the DL-LCD scheme,and the trajectory optimization module adopts the classical g2o pose graph scheme.DL-SLAM can estimate the monocular scale online,detect the loop-closure path online,and optimize the camera pose online.The camera trajectory is updated online,and it has stable performance in any scene environment and finally recovers the incremental camera motion trajectory. |
