GPU-SIFT based real time binocular vision positioning method

Abstract

The invention relates to a GPU-SIFT based real time binocular vision positioning method. The method comprises the following steps: 1) adopting a parallel binocular camera for acquiring a stereo image video of left and right eye images in a moving process of a robot or a mobile platform; 2) adopting feature point matching for acquiring corresponding matching points in front and rear frames of a shot video in a motion process; 3) solving a motion equation by changing the matching points in an imaging space coordinate or establishing a three-dimensional coordinate, thereby acquiring the displacement of the camera by estimation; 4) acquiring the position and rotation angle at each moment of the camera in advancing and combining with kalman filtering, thereby acquiring the advancing path of the camera in the whole process and realizing real-time binocular vision positioning on the robot or the mobile platform. According to the invention, GPU-SIFT is adopted for accelerating the SIFT feature matching process, the binocular vision positioning is adopted for realizing the real-time binocular vision positioning of the robot or the mobile platform, the higher positioning accuracy can be acquired and the extendibility, practicability and environment applicability are higher.

Description

[0001] Technical field: [0002] The invention relates to the technical field of robot vision positioning and navigation, in particular to a real-time binocular vision odometer system based on GPU-SIFT. [0003] Background technique: [0004] With the continuous development of robots and computer vision, cameras are increasingly used in the visual positioning and navigation of robots. Robot positioning mainly includes code disc, sonar, IMU, GPS, Beidou, laser scanner, RGBD camera and binocular camera for visual positioning. The stroke is deduced to achieve positioning, but this positioning method has a large error and inaccurate positioning in the case of sand, grass or wheel slippage. Sonar positioning mainly relies on ultrasonic sensor emission and return signal analysis to determine obstacles for positioning and navigation, but the resolution of sonar is low, and there are more noises in the signal, which easily interferes with positioning. Robots using IMU for positioning...

AI Technical Summary

Problems solved by technology

However, SIFT feature extraction and description consume a lot of time and it is difficult to achieve real-time image matching. GPU-SIFT, which uses GPU to accelerate the process of SIFT feature extraction, description and matching, can greatly accelerate the SIFT feature matching process and realize real-time SIFT feature matching.

Images