CN110070577B - Visual SLAM key frame and feature point selection method based on feature point distribution - Google Patents

Abstract

The invention discloses a visual SLAM key frame and feature point selection method based on feature distribution, and relates to the field of visual synchronous positioning and mapping. Specifically, the invention provides a method for selecting key frames and feature points during visual synchronous positioning and mapping. When there are more ranges not stored in the map at the edge part of a certain frame image, the frame is defined as a key frame. The selected characteristic points should cover the whole image as much as possible in a certain frame of image. The invention improves the tracking stability, reduces the number of key frames and key points and meets the engineering application requirements while not affecting the visual synchronous positioning and mapping precision.

Description

Visual SLAM key frame and feature point selection method based on feature point distribution

Technical Field

The invention belongs to the technical field of visual synchronous positioning and mapping, and relates to a visual SLAM key frame and feature point selection method based on feature point distribution.

Background

The research of Visual synchronous positioning and mapping (Visual SLAM) is a big hot spot in recent years, the traditional SLAM mainly uses a laser radar as a sensor, but with the development of computer vision correlation, the laser radar is used in unmanned aerial vehicle,

Visual localization and mapping is also beginning to be used on mobile robots, AR wearable devices. The visual positioning and mapping technology is one of core technologies for supporting unmanned aerial vehicles, mobile robots and other equipment to realize autonomous movement.

The visual positioning and mapping technology mainly comprises the steps of preprocessing sensor data, front-end visual odometer, rear-end nonlinear optimization, loop detection, mapping and the like. The sensor data mainly comprises an inertial measurement unit, a camera and the like. Common cameras include monocular cameras, binocular cameras, RGB-D cameras, and the like. The monocular camera has the advantages of low cost, no limitation of distance, but the defects of scale uncertainty and the like, the binocular camera can calculate depth, the distance is also not limited, the configuration is complex, and the calculated amount is large. The RGB-D camera can actively measure depth, has good reconstruction effect, but has smaller measurement range. The front-end visual odometer is used for quantitatively estimating the movement of the camera according to the images of the adjacent frames, if the movement tracks of the adjacent frames are integrated, the movement track of the camera can be deduced, positioning is realized, the position of the camera at each moment is estimated, and the position of the pixels in space is obtained, so that the map is obtained. The front-end visual odometer is the key of accurate positioning and image construction. Because the front-end visual odometer only estimates the motion of adjacent frames, a larger track drift can be generated after a certain time of accumulation due to errors generated in each estimation. To reduce track drift, back-end optimization and loop-back detection are required. The rear-end optimization is performed according to the result obtained by the front-end visual odometer, so as to obtain the best pose estimation. The back-end optimization mainly comprises two methods, namely optimization based on a filtering theory and nonlinear optimization (graph optimization). Because SLAM systems are nonlinear non-Gaussian systems, graph optimization has become the dominant approach, and traditional filtering theory uses less and less. Loop detection is also a method for reducing drift, and by judging the similarity of images, it is confirmed whether the current position is reached, so that the accumulated error is eliminated from the new adjustment map and track. The synchronous positioning and mapping technology can establish different maps according to the needs, and common maps include a 2D grid map, a 2D topological map, a 3D point cloud map and the like. The 2D grid map is most commonly used, a map can be obtained to obtain information of a two-dimensional plane, the 2D topological map emphasizes the communication information among elements, many details are removed, and the expression is more refined. The 3D point cloud map can be used for visual reconstruction of a real scene, but the occupied space is too large, and redundant information is too much.

All parts of visual positioning and image construction need to be mutually connected and closely matched, wherein a front-end visual odometer is a key ring for accurately positioning and image construction. The existing method has the defects of up to standard precision, insufficient quantity of matched characteristic points, insufficient tracking stability and difficulty in being applied to engineering.

Disclosure of Invention

The invention aims at: the method aims to solve the problems that the existing algorithm is not stable enough in tracking and redundant information stored in a map is too much. In particular, the key frame judgment and the feature point selection are required to achieve a certain precision in engineering practice, and meanwhile, stable tracking is required. If the environment map is used for visual positioning and navigation of the robot, more characteristic points must be matched to realize tracking stability. The invention provides a visual SLAM key frame and feature point selection method based on feature point distribution.

The technical scheme of the invention is as follows:

as shown in fig. 1, the visual SLAM key frame and feature point selection method based on feature point distribution is characterized by comprising the following steps:

s1: inputting a certain frame of image obtained by a camera;

s2: extracting the frame characteristic points, and matching with the latest (last) key frame to obtain matching points;

s3: judging whether the pose amplitude of the camera is changed too much, if so, executing S7; no, dividing the current frame into a plurality of areas a ₁ ,a ₂ ,…，a _n The method comprises the steps of carrying out a first treatment on the surface of the The partitioning method includes but is not limited to the method shown in fig. 2;

s4: counting the number of feature points of each region;

s5: and according to the result of S4, counting the number of the effective areas of the whole graph. The effective area is an area with the number of the characteristic points larger than a threshold value;

s6: judging whether the number of the effective areas without the matching points is not less than 2, if yes, executing S7; if not, the frame is a non-key frame and ends;

s7: selecting the frame as a key frame;

s8: handle a ₁ Setting the current area;

s9: is the current region feature points? If yes, executing S10; if not, setting the next area as the current area;

s10: selecting a feature point with the minimum depth of the current area, storing the feature point into a map, and removing the feature point from the current area;

s11: judging whether the feature points are stored in the map completely or the number of the feature points stored in the map reaches a threshold value, if yes, finishing the storage of the key frame information; if not, executing S12;

s12: setting the next area as the current area, a _n The next region of a is a ₁ S9 is performed.

The method has the beneficial effects that the prior method mainly aims at improving the precision when the mobile robot performs visual positioning and mapping, so that the related algorithm cannot keep the tracking stability and cannot be effectively put into practical use in engineering. The invention can select more representative key frames and more quality characteristic points under the condition that the precision is basically equal to the effect of the existing algorithm, so that the key frames required to be selected are obviously reduced, but the number of the matched characteristic points can be obviously increased. The invention improves the tracking quality, reduces the occupation of the storage space, and can be better put into engineering use compared with the existing algorithm.

Drawings

FIG. 1 is a flow chart of the present invention;

FIG. 2 is one possible implementation of the region division method used in the present invention;

FIG. 3 is a graph comparing the number of key frames obtained by performing experiments on EuRoC data sets in accordance with the present invention;

fig. 4 is a graph comparing the number of matching feature points obtained by experiments performed on the EuRoC dataset according to the present invention.

Detailed Description

The embodiments of the present invention have been described in detail in the summary section, and are not further described herein.

Fig. 3 and 4 are diagrams showing the comparison of the method of the present invention and the conventional method under the same condition, it can be seen that the number of key frames required by the method of the present invention is significantly reduced, and the number of matched feature points is significantly increased. The required storage space is reduced, and meanwhile, the tracking quality is improved. Embodying the utility of the invention.

Claims (1)

1. The visual SLAM key frame and feature point selection method based on feature point distribution is characterized by comprising the following steps of:

s1: inputting a certain frame of image obtained by a camera;

s2: extracting the characteristic points of the frame image, and matching with the previous key frame to obtain matching points;

s3: judging whether the pose amplitude of the camera is changed too much, if so, executing S7; no, dividing the current frame into a plurality of areas a ₁ ,a ₂ ,…，a _n ；

S4: counting the number of feature points of each region;

s5: according to the result of S4, counting the number of the effective areas of the whole graph, wherein the effective areas refer to areas with the number of the characteristic points larger than a threshold value;

s6: judging whether the number of the effective areas without the matching points is not less than 2, if yes, executing S7; if not, marking the frame as a non-key frame, and ending;

s7: selecting the frame as a key frame;

s8: handle a ₁ Setting the current area;

s9: judging whether the current area has characteristic points or not, if so, executing S10; if not, setting the next area as the current area;

s10: selecting a feature point with the minimum depth of the current area, storing the feature point into a map, and removing the feature point with the minimum depth from the current area;

s11: judging whether the feature points are stored in the map completely or the number of the feature points stored in the map reaches a threshold value, if yes, finishing the storage of the key frame information; if not, executing S12;

s12: setting the next area as the current area, a _n The next region of a is a ₁ S9 is performed.

Images