Algorithms Based On RGB-D And Monocular Vision For Simultaneous Localization And Mapping

With the development of science technology,computer network and hardware,simultaneous localization and mapping(SLAM)has become one of the focus in the field of intelligence mobile robots undoubtedly.It plays a crucial role in the autonomous movement of robots.SLAM can be implemented using many methods,in general can be divided into two types: one is based on filter and the other is based on pose-graph optimization.For the use of image information as data source SLAM problem is also known as visual SLAM(VSLAM).In the dynamic,complexity and large scale environment,using vision as the only external perception of information to solve the SLAM problem is still an active research field.As for this,the main contents of this paper are as follows:Firstly,the advantages and disadvantages of different visual sensors are compared and the RGB-D mapping algorithm based on depth camera is constructed with reference to the “pose-graph optimization”.Aiming at the problem of weak robustness and real-time and poor matching accuracy of traditional visual features,an RGB-D mapping algorithm based on ORB visual feature is proposed.Then,the different visual features(ORB?SIFT,SURF,FAST,GridFAST)are applied into the RGB-D mapping algorithm and compared their influence on the real-time,accuracy and relolization capability of the whole mapping algorithm.As the experiment shows,ORB feature is better than the others in aspects of robustness,real-time and matching accuracy and the RGB-D mapping algorithm based on ORB feature is the best in real-time,mapping accuracy and relocalization accuracy.Secondly,an improved keyframe selection algorithm is proposed for the problem that the traditional keyframe selection algorithm is simple and the number of keyframes in the whole SLAM process is increased sharply.Based on this,the RGB-D SLAM algorithm based on ORB feature is constructed.The improved keyframe selection algorithm not only integrates the distance of relative motion between frames,the feature points tracking and minimum visual change to select the keyframe,but also checks redundant keyframes and deletes them.The experiments on the RGB-D dataset show that the improved keyframes algorithm can select keyframes more accurrately and timely.Reducing the redundant keyframes of RGB-D SLAM algorithm and improvingthe real-time and localization accuracy at the same time.Finally,for the use of RGB-D camera with less flexibility and the feature method with poor robustness,the direct method is used to realize the Monocular SLAM algorithm.The Algorithm can be used in both indoor and outdoor environment with the common monocular camera as sensor to obtain environment information,which overcomes the limitations that depth camera can only be used in the indoor environment.Compared with feature method,direct method deals with the pixels gray of the images directly,which can make full use of the environment rich information and make good estimation as well when salient features are not obvious.Experiments show that direct method works well when features are not obvious and can locate and map quickly and accurately no matter in the indoor or outdoor environment.