Calculation method of monocular visual odometry based on deep learning and attention mechanism

Abstract

The invention discloses a monocular visual mileage calculation method based on deep learning and attention mechanism. The specific steps include: first collecting the original monocular color image, and cutting the image size to a uniform size; Input the i+1th picture into the PWCnet optical flow calculation module to obtain the optical flow vector field, and divide the optical flow vector field into 4 optical flow sub-vector fields according to 4 quadrants; input the 4 optical flow sub-vector fields The convolutional neural network obtains 4 sub-eigenvectors respectively; merges the 4 sub-eigenvectors into a total eigenvector, and then inputs the total eigenvector into a fully connected network to obtain an estimated pose vector; collects pictures in real time and sends them to convolution in turn The neural network obtains several consecutive estimated pose vectors, and obtains the estimated mileage through several consecutive estimated pose vectors. The invention has good visual mileage calculation accuracy and calculation speed.

Description

technical field [0001] The invention belongs to the field of autonomous positioning of mobile robots, and in particular relates to a calculation method of a monocular visual odometer based on deep learning and an attention mechanism. Background technique [0002] Visual odometry technology is a front-end technology in simultaneous visual positioning and map construction. The inter-frame pose estimation obtained by the visual odometry can obtain a local map, and the local map can obtain a global map of the path traveled by the odometry after the back-end optimization. In this way, tasks such as map construction and 3D reconstruction can be further carried out. [0003] Visual odometry technology is widely used in mobile robot autonomous positioning, automatic driving, and virtual reality technology, and it is a hot research field in recent years. The main research task of visual odometry technology is to use visual feature vectors for accurate inter-frame pose estimation. ...

AI Technical Summary

Problems solved by technology

Since the design of eigenvectors is extremely artificial, the designed eigenvectors have limitations, and the use of eigenvector points ignores other information except eigenvector points

And the camera may move to places where feature vector points are missing, where there is no obvious texture information

In addition, the extraction of feature vector points and the calculation of descriptors are time-consuming.

The direct method estimates the camera motion and the spatial position of the pixel by minimizing the photometric error, which can achieve better results in scenes where the feature vector is not obvious, such as corridors or smooth walls, but it is only applicable to the range of motion Smaller and the overall brightness of the picture does not change much

[0004] The traditional calculation method of visual odometry has the following two problems: first, it is necessary to know the internal parameters of the camera; second, there is a difficult trade-off between accuracy and speed

Images