A visual slam method and device based on dynamic target detection

Abstract

The invention discloses a visual SLAM method based on dynamic target detection, which uses the target detection network Yolov3 to temporarily remove the potential dynamic area of ​​the image, optimizes the homography matrix through re-projection errors, solves the motion compensation frame and obtains a four-frame difference map, and then performs The four-frame difference image is filtered, binarized and morphologically processed. At the same time, the Yolov3 network is combined to optimize the dynamic target detection results, so as to obtain the improved dynamic target area. Finally, the feature points of the static area are used for visual SLAM tracking. , Mapping and loop detection. The above method uses the deep learning target detection network to first eliminate the potential dynamic area in the scene, roughly estimates a homography matrix, and judges the feature points on the potential dynamic area based on the combination of reprojection error and inter-class variance. It is used for the calculation of the homography matrix to optimize the homography matrix, thereby improving the accuracy of the homography matrix.

Description

technical field [0001] The invention belongs to the technical field of image processing, and in particular relates to a visual SLAM method and device based on dynamic target detection. Background technique [0002] Simultaneous positioning and map construction (SLAM, Simultaneous localization and mapping) technology is more and more widely used in the fields of robot positioning and unmanned driving. Among them, the vision sensor has the characteristics of portability and low cost, so it is widely used in In SLAM technology, most of the traditional visual SLAM algorithms assume that the camera is in a static environment, such as Orbslam2, DSO, SVO, etc. When there are dynamic areas in the scene, the feature points extracted by visual SLAM on dynamic objects will affect the accuracy of the algorithm. . [0003] To solve the problem of decreased accuracy of visual odometry in dynamic scenes, the commonly used method is to detect dynamic objects on the image first, remove the ...

AI Technical Summary

Problems solved by technology

[0002] Simultaneous positioning and map construction (SLAM, Simultaneous localization and mapping) technology is more and more widely used in the fields of robot positioning and unmanned driving. Among them, the vision sensor has the characteristics of portability and low cost, so it is widely used in In SLAM technology, most of the traditional visual SLAM algorithms assume that the camera is in a static environment, such as Orbslam2, DSO, SVO, etc. When there are dynamic areas in the scene, the feature points extracted by visual SLAM on dynamic objects will affect the accuracy of the algorithm.

[0003] To solve the problem of decreased accuracy of visual odometry in dynamic scenes, the commonly used method is to detect dynamic objects on the image first, remove the feature points in the dynamic area, and retain the feature points in the static area for visual SLAM tracking and mapping. In images with a large area ratio, the removal of dynamic areas will greatly affect the accuracy of visual SLAM tracking and mapping

[0004] The defects in the prior art are mainly caused by the following reasons: using the deep learning target detection network alone, it is possible to pre-classify movable objects such as people and cars as potential dynamic targets, but it is impossible to judge whether the potential dynamic targets are In a real state of motion, if the potential dynamic target is in a stationary state, too many stationary feature points may be removed

Algorithms that need to combine depth information for dynamic detection may lead to inaccurate classification when the depth information of some areas of the image is uncertain, or when the depth of the foreground and foreground are relatively close

Images