Research On Key Technologies Of Sustainable Online Simultaneous Localization And Mapping With Machine Vision

Vision based simultaneous localization and mapping(v SLAM)refers to the set of algorithms and tools that realize the estimation of the state of a robot equipped with visual sensors only without any heuristic knowledge while constructing the environment map simultaneously.The estimation on camera movement and pose are based on continuous video frames during the camera movement,and the mapping process is based on surrounding information contained in the perceived environment.v SLAM is recognized as the underlying algorithms which support the autonomous robots and augmented reality(AR).It receives a great deal of attention in industry and academic sectors due to its wide range of applications and high academic value.The function environment of traditional v SLAM is small in scale and contains very few dynamic factors,where there is no obvious illuminate variation.However,the actual function environment of autonomous robot is usually complex and large,and different sensors perceive the environment in different way.Specifically,one scenario may contain different image features under the illumination of various strengths and positions.There are also a lot of overlap and redundant texture features in the continuous video frames captured by cameras.One single scenario may have different perceptions with multiple sensors,and the data size increase exponentially with scene expansion and mapping time elapses.The continuous video frames contain a lot of similar texture features and traditional approaches can not perform place recognition effectively.This dissertation undertakes relevant researches on problems of illumination coherence and complex texture features,fast increase of map storage with scene expansion and scene ambiguity issue in loop closure detection of v SLAM systems.The main contributions can be summarized as follows:1.Targeting the problems of illumination consistency and complex textures in augmented reality application of v SLAM,a filtering algorithm based on Laplacian of Gaussian(Lo G)operator is proposed.First,edges and details are intensified in AR scenario by filtering streaming image pyramids with Lo G operator.Then the zero mean sum of squared difference(ZMSSD)threshold is set according to the values of average of final accept score(AFAS)and average search length(ASL)to adapt to the frames contrast variation after Lo G transformation.A bilateral optical flow method is implemented for improved matching precision.Finally,an deep learning pipeline is implemented to guide the feature matching for improved accuracy in relocalisation,which results in accurate camera pose recovery in case of tracking failure.Experiments demonstrate that the proposed system improves the illumination variation tolerance and the virtual objects can register in 3D environment precisely in AR application,which keeps robust performance under various scenarios.2.The traditional environment modeling approach is unable to merge data from various sensors and the working diameter of depth camera is also limited.To address the problems,an environment perception algorithm based on probabilistic octree is proposed.The algorithm is able to uniformly express the spare,semi-dense and dense point cloud models from different sensors.Benefiting from the pruning and merging strategies of octree,the proposed algorithm reduces the storage requirement and extends multi-granularity representation of environment models.Upon environment change caused by sensor movement,the three kinds of model are fused with extended Kalman filter(EKF)and a temporal fusion probabilistic octree model is constructed.Experiments demonstrate that the v SLAM system estimates the camera pose precisely with the constructed environment models.The time and storage requirements in environment modeling are also reduced,which facilitate the migration of the algorithm to embedded device and function for long-lasting time.The converted map provides the passable information in the physical environment for v SLAM application.The generated map can be applied to navigation and obstacle avoidance purposes in various applications.3.Aiming at overcoming the scene ambiguity problem and undetectable issue of bidirectional loops,a Bo VW-based loop closure detection approach via maximization of mutual information(MMI)mechanism is proposed.The proposed algorithm aims at improving the quality of visual vocabulary by extracting visual words of high discriminative power.It recognizes the vocabulary construction process as data compression.The discriminative visual words are extracted through the optimization of objective function,which finds the balance between data compression and relevant information preservation.The proposed algorithm has potentials in applications to content based image retrieval(CBIR),image matching and image classification.Experiments demonstrate that the improved loop closure detection not only reduces the detection of false positive loops,but also makes the detection of bidirectional loops possible.The improvement helps in reducing the accumulated error in long term pose estimation and mapping,which expands the application range of v SLAM system.