Self-positioning In Dynamic Environment And Loop-closure Detection For Mobile Robot

Localization and perception are the basic problems to be solved for the autonomous robot,and the Simultaneous Localization and Mapping(SLAM)is the solution to this problems.Because a camera has many advantages such as low cost,light weight,and rich information acquisition,the Visual SLAM has gradually become a hotspot in robot research.An important function of Visual SLAM is to estimate the ego-motion of camera through the image acquired by the camera.Therefore,this disseration focuses on two methods for localization in the visual SLAM system: visual odometry and loop closure detection.Recently,the research on visual odometry is mainly on the basis that the localization of camera is calculated on the assumptions that the majority of the scene is stationary.If there are moving objects in the environment,there will be a large error in the camera selfpositioning.A novel feature segmentation based stereo visual odometry is proposed in this disseration,aiming at the self-positioning problem in dynamic environment.The proposed method extracts and matches the features from the current frame and keyframe images,and then calculates the transformation according to the matched features.In order to eliminate the disturbance of moving objects in the environment,the features are divided into static and dynamic components based on the epipolar constraint.Finally,only the matched static features are used to estimate the ego-motion of the camera,and the self-positioning of the camera in the dynamic environment is achieved.Loop closure detection is a critical part of the Visual SLAM system,which can reduce the accumulating drift of visual odometry.This disseration proposes a loop closure detection algorithm based on convolutional neural network(CNN).Different from the previous work,the high-dimensional feature vector generated by model is directly used for loop detection.After extracting the features of keyframes from the CNNs model,the feature vector is preprocessed by reducing the dimension and binarization,which greatly improving the calculation speed when detecting the closure.The experimental results show that this algorithm can greatly improve the speed of feature vector matching while guaranteeing the recall and accuracy in comparison with the traditional methods.