Research On Visual SLAM Systems For Dynamic Scenes

At present,artificial intelligence is playing an important role in the rapid development of science and technology.The intelligent robot which is one of the products of artificial intelligence,has been widely used in many scenes,such as storage and transportation in the warehouse,medical treatment,rescue in the wild,service robots and so on.Visual simultaneous localization and mapping(SLAM)is a key technology in the research of the intelligent robot.Because of the assumption that the environmrnt around robots is static,dynamic objects existing in the scene will cause the decline in the robustness and accuracy of most SLAM systems.This paper proposes a visual SLAM system for dynamic scenes aiming at decreasing the bad effects of dynamic objects.It does research on the visual odometry,loop closure detection and the keyframe detection,which are the parts of the SLAM system.The main contents of this paper are as follows:1)Firstly,the bad effect of moving objects on the visual odometry is introduced according to the feature extraction and matching.Then,an improved dynamic and static feature point division scheme based on geometric constraints is added into the visual odometry framework of ORB-SLAM2.When the dynamic points are detecting,they will be filtered and only the static points are used to estimate the camera pose.Experiments show that the visual odometry proposed has good performance in the dynamic scenes.With more quantity feature points retained,the possibility of the failure in pose estimation is getting smaller.2)Secondly,in order to avoid the instability and failure of the loop closure detection in dynamic scenes,the static background area will be compensated when the dynamic points are filited accounding to the moving object detection results.Then the static feature points are projected into the visual dictionary tree to obtain the description vector of the image.In addition,the stable spatial relationship between visual words is taken into considering to get the visual phrase and visual phrase set.The representation of the TF-IDF and the similarity calculation function are also changed,What’ more,the visual words and visual phrases are used jointly to calculate the similarity between images.Experiments show that the proposed loop closure detection algorithm can effectively reduce the perceptual ambiguity and improve the recall rate.3)Finally,this paper briefly introduces the four threads of the SLAM system.Besides,considering the importance of the keyframe in the whole system,a proportional and differential controller is used to optimize the selecting sheme of keyframes.This paper also gives a brief introduction to adjustments in the whole system.Experimental results show that the proposed SLAM system in this paper is more robust and accurate than other previous SLAM systems in dynamic scenes.