Research On Visual SLAM In Dynamic Scenes Based On Unsupervised Learning Of Optical Flow Estimation

Visual sensors based simultaneous localization and mapping technology(Visual SLAM)is a key technology esof mobile robots.It enables robots to estimate its own motion and build environment models in an unknown environment,which is an important prerequisite for in-depth research and application of mobile robots.The current traditional geometric-based visual SLAM method restricts the working environment to a static scene,and when disturbed by dynamic objects,there will be a great positioning drift,as well as smearing and distortion of the mapping;visual SLAM method based on semantic information eliminates the interference of common dynamic objects such as pedestrians and vehicles by training a neural network to detect and remove dynamic objects,but the types of objects that can be processed are limited by the manual annotated data set.In order to enhance the robot's ability to estimate its own pose and perceive the external environment in dynamic scenes,a dynamic scene visual SLAM method based on unsupervised learning optical flow estimation is studied in this paper.The main research work is as follows:First,aiming at the problem of dense optical flow estimation of dynamic scenes,photometric consistency and smoothness consistency constraints of the image are used to construct an unsupervised learning network for optical flow estimation.At the same time,epipolar geometric constraints are introduced to improve the model's optical flow estimation accuracy in dynamic scenes.On this basis,data augmentation is incorporated into the unsupervised learning process,which improves the generalization ability of the model in complex scenes such as dynamics,insufficient lighting,and motion blur.An accurate optical flow estimation for dynamic scenes is achieved,which provides important priors for subsequent motion segmentation,localization and mapping research.Secondly,aiming at the problem of accurate localization and mapping in dynamic scenes,results of optical flow estimation are used to perform image segmentation,which is combined with visual and inertial information to perform accurate motion segmentation on the image.The dynamic object mask obtained by motion segmentation is used to remove the dynamic visual features,and the static features are tightly coupled with the IMU information.Besides,loop detection and optimization are performed to realize an accurate estimation of the camera pose.Using the results of pose estimation,optical flow estimation,and motion segmentation,the dense depth of the environment is calculated,and a dense map is constructed where dynamic objects are removed.Finally,the proposed optical flow estimation method and the optical flow-based visual SLAM method are verified on benchmark datasets and real-world scene data,and are compared with several advanced methods.Experimental results show that the optical flow estimation method proposed in this paper get a leap performance of accuracy,time effiency and generalization ability.The visual SLAM method based on optical flow can accurately and estimate camera pose and construct dense environment maps in dynamic scenes in real-time.Compared with several advanced algorithms,our method has significant advantages.