Real scale obtaining method of monocular vision odometer

Abstract

The invention relates to a real scale obtaining method of a monocular vision odometer. The method comprises: (1), a probability of belonging to a pavement of each pixel point in each frame of image shot by a monocular camera is obtained; (2), a rotation matrix R and a camera relative displacement vector of camera motion between two adjacent frames of images shot by the monocular camera are obtained; (3), a pavement model being a plane approximately is constructed and a position of a point, on the pavement in a next frame of image, in a previous frame of image is obtained by using a homography method; (4), a weighted average value of a difference between pixel values of corresponding pixel points in the two adjacent frames of images is calculated and a corresponding pavement model parameter on the condition of minimizing the weighted average value of the difference of the pixel values is calculated; and (5), an absolute scale is obtained based on the actual heights of the camera and the height in the pavement model, and an actual displacement vector of the camera is calculated. Compared with the prior art, the method has advantages of low cost, simple operation, high accuracy and robustness, and high efficiency and the like.

Description

technical field [0001] The invention relates to a scale acquisition method, in particular to a real scale acquisition method of a monocular visual odometer. Background technique [0002] Visual odometry can obtain the trajectory of smart cars only based on visual image information, which is of great value in autonomous positioning and navigation of smart cars. Compared with GPS, inertial navigation and code disc with large errors, visual odometer can have higher accuracy; compared with expensive and bulky laser sensors, visual information can be obtained conveniently and at low cost. Monocular visual odometry is more suitable for large-scale scenes such as outdoors than binocular visual odometry. However, due to the loss of depth information in the monocular visual odometry, scale information cannot be obtained directly. [0003] At present, additional sensors are generally required to restore the monocular visual scale, such as VIO combined with vision and inertial naviga...

AI Technical Summary

Problems solved by technology

However, due to the loss of depth information in the monocular visual odometer, the scale information cannot be obtained directly

[0003] At present, additional sensors are generally required to restore the monocular visual scale, such as VIO combined with vision and inertial navigation, but the most ideal and convenient way is naturally to restore it with a purely visual method, which is generally based on road surface information. , there are currently two methods. One is to assume that a part of the lower middle area of ​​the image is the road surface, and then use this area to realize the homography method to obtain the road surface model, but this assumption is not always true. When the vehicle is turning Or when there is a vehicle blocking in front, most of this area may not be the road surface, and the estimated model at this time must be wrong

Another method is to use the triangulation method to calculate the spatial position of the matched feature points after obtaining the R, t without scale, and then use the Ransac method to estimate the road surface model. After estimating the initial model, it can be trained according to the color of the road surface A classifier is used to estimate the probability that the next pixel point is the road surface, and the points with high probability are selected to find the road surface model. The disadvantage of this method is that because the general road surface area is flat, there will be fewer feature points, and Ransac may not be able to estimate the road surface model. Accurate results are obtained, and in the process of feature point matching tracking, triangulation, and Ransac model estimation, the algorithm will be relatively slow, and it is difficult to achieve real-time performance

Images