The Use Of Object Tracking In Vision SLAM

In recent years,thanks to the development of optical equipment,visual sensors are widely used in mobile robots and other related fields.Compared with other non-visual sensors such as laser,sonar and so on,the visual-based way can obtain relatively more intuitive and rich information,and can give a higher dimension of the pose description,so the visual positioning and building(Visual SLAM)gradually Has become one of the hotspots in the field of mobile robots.The current research of visual SLAM can be divided into indirect method and direct method.The main idea of the indirect method is to use sparse features to perform scene tracking and pose estimation.The existing research shows that this method can solve the problem of Bundle Adjustment(BA)by using the point feature,and it has a high degree of robustness to the point feature extraction and has a large computation.The direct method utilizes the pixel information to optimize the motion and the motion estimation by optimizing the measurement error between frames.Relative to the indirect method,the direct method avoids the detection,description and matching of feature points,and can provide semi-sparse or dense point cloud structure.Under the influence of this research background,recent studies have begun to solve SLAM problems in a direct or semi-direct way.But the direct method relies to some degree of image correction,and the relatively large camera movement will affect the photometric error,thus affecting the effect of scene reconstruction.Based on the above discussion,we found that in the recent study,the scene tracking and pose estimation and follow-up work are more interdependent.To some extent,the tracking method determines the validity of localization and mapping.The research idea of this topic is to solve the scene and the back-end independent solution,so as to effectively avoid the excessive dependence on the feature points,while improving the speed of reconstruction,as follows:Firstly,it is proposed to use the tracking method of kernel function to carry out independent scene tracking.The kernel-based method has a good classification effect.and it has a good classification ability for regions with small gradient or small texture.Therefore,the region of interest is used as the local continuous tracking of the scene to be achieved,that is,the local information of the scene is used as the landmark,And then to achieve the scene of the local track division.This way on the one hand can effectively deal with the problem of tracking loss,on the other hand can provide a non-global description of the scene means.Secondly,the research idea provides a more adaptable application framework for the visual SLAM backend,and the traditional indirect and direct methods can be included in the framework.Specifically,through independent local scene tracking,point features can be used to solve SLAM back-end.Similarly,the direct method based on inter frame photometric error can also take our scheme.Finally,this topic has been verified by a semi-direct method.