Research On Visual Slam Based On Deep Learning And Edge Detection In Dynamic Scenes

Simultaneous Localization and Mapping is key point when the robot executing intelligent task under complicated environment,it's also the research hot spot of robot and unmanned now.Building map and locating itself in the unfamiliar environment,all of these can be done with sensor only when SLAM was used by robot.With the continuous development of slam research,it's possible to do the real time location and map building under the large-scale static scene.But there are still two shortages: the first is when it's under the dynamic environment,the precision and robustness of the location can be easily impacted by the dynamic objects in the scene,they can disturb the feature matching which is calculated in the procedure of pose estimation;the second is low efficiency and poor readability of the 3D map building,especially in the large-scale scene it cannot guarantee the real time,and the map contains ghosting.For the problem above,an improved system based on the traditional SLAM is proposed by this paper,which aims to promote the precision of the location and map building under dynamic environment,keep the real time at the same time.The principle of this optimization is as follows:(1)Procedure of dynamic feature points filtering is added during pose estimation.The semantic and optical flow information is integrated to filter dynamic feature points,it is proposed by this paper for moving objects processing under dynamic environment.Mainstream semantic segmentation and object detection algorithms based on deep learning is researched and compared on their accuracy,precision and efficiency.After comprehensive assessment,object detection classifier was integrated to ORB-SLAM structure and sparse LK optical flow information is applied to estimate dynamic objects,using the static feature points only to match and project,construct and minimize the reprojection error to get pose.(2)Dynamic objects elimination and key frame selection optimization is done during point cloud map building.The bigger tracking error is,the bigger world coordinates error is,and the ghosting is introduced.Dynamic object contained in the scene contributes to poor readability of map,so the edge of them was detected by semantic and optical flow information and pixel of them was filled,then combined pixel with common visual feature points to estimate the effectiveness and redundancy to get the key frame for drawing,this is good for simplifying the key frames which helped to decrease the calculation amount brought by large scale points.(3)Performance of the system proposed by this paper is tested.The dynamic objects detection,tracking,location and map building is experimented with TUM datasets.After experiment and comparing with ORB-SLAM,DS-SLAM,Dyna SLAM,the system implemented by this paper can run better under high dynamic environment,both static and dynamic objects can be distinguished correctly,the pixel of dynamic objects can be eliminated accurately,it achieves the accurate positioning and 3D map building,which proved that the system proposed by this paper has better precision and robustness.