Research On Optimization Method Of Visual Odometer In Dynamic Scene

Visual odometry is to estimate the ego-motion of an agent using sequence images.It is an important topic in computer vision and broadly used in autonomous robot technology,such as positioning and navigation.With the rapid development of mobile autonomous robots,robots' various automation tasks,such as automatic driving,search and rescue,etc.,put forward higher requirements for the positioning of mobile robots.Most visual odometry methods have achieved good results in static scenes,but still fail or produce large error in pose estimation in dynamic scenes or static scenes containing moving objects,such as animals,pedestrians,vehicles,etc.This paper deeply studies the principle of visual odometry method and the general flow of current mainstream visual odometry method.For the problems existing in dynamic scenes,the feature point matching,local map management and pose estimation in the algorithm are optimized,Improves the pose estimation accuracy of visual odometry in dynamic scenes.The main work and research contents of this paper are as follows:(1)A method for pre-detecting moving object feature points based on Inter-frame feature point moving distance histogram is proposed.At the same time,a key frame selection and deletion strategy in dynamic scene is proposed,Then,a local map construction and management method including virtual feature points in dynamic scene is proposed,which reduces the interference of moving objects in dynamic scenes.By distinguishing the feature points on the static scene and the moving object,and generating the virtual feature points according to the uniform motion model of the feature points on the moving object,the local map is added for the pose estimation,aiming at improving the accuracy of the visual odometer in the dynamic scene.Experiments show that the key frame management strategy proposed and the local map construction and management method based on virtual feature points in this paper effectively improve the accuracy of visual odometer pose estimation under dynamic scenes.(2)The local map construction and management method including virtual feature points in the dynamic scene is applied to the visual odometry,and a complete visual odometry method suitable for dynamic scenes is designed and implemented.At the same time,aiming at the matching step of feature points in the process of visual odometry,a feature point matching and filtering method based on camera uniform motion model is proposed.For the uncertainty of virtual feature points,the pose of the camera is calculated by RANSAC PnP method as the initial value.Then a new re-projection error function is proposed to iteratively optimize thecamera pose to obtain more accurate camera pose estimation.The experimental comparison shows that the proposed method achieves good results in both low-level dynamic scenes and high-level dynamic scenes.Especially in high-level dynamic scenes,the proposed method effectively reduces the interference of moving objects on the pose estimation of the algorithm,and achieves a relatively large effect.(3)The visual odometry pose recovery and evaluation system was designed and implemented.The system consists of three functional modules:data selection module,result display module and performance evaluation module,which are used for algorithm data set selection and parameter setting,algorithm execution result display and quantitative evaluation of algorithm execution results.The system can visually show the effect of the visual odometry,which is convenient for comparison with the performance of various existing visual odometry methods.