基于 RGB-D 相机的 SLAM 技术研究
- II -
Study on SLAM Based on RGB-D Camera of Mobile Robot
Abstract
With the development of science and technology, mobile robots have gradually entered
people's lives and become a very important part. At present, the requirements for robots in
various fields include the autonomous positioning and path planning of robots. To achieve
these functions, we must first realize the simultaneous localization and mapping of the robot,
namely SLAM. The system of SLAM can complete the robot's autonomous positioning and
environment map construction through the environmental information obtained by the sensor
when the environment of the robot is unknown.
This paper studies the visual SLAM system based on RGB-D camera. The main research
work is as follows.
(1) A classic framework of visual SLAM is introduced, then is the complete circuit of the
SLAM system based on KINECT. And the structure, principle and imaging mode of KINECT
is also introduced. The KINECT is calibrated so that the internal and external parameters of the
camera are obtained.
(2) The realization of the visual odometer at the front of the SLAM system is based on the
method of feature. The image features are selected as ORB (Oriented FAST and Rotated
BRIEF) features. Scale invariance and rotation invariance are added to the ORB feature by
image pyramid and gray centroid method. Then, the Hamming distance is used as the metric for
feature matching. To eliminate the mismatching in the matching result, the distance threshold is
set and the suboptimal matching is used. The experimental results show that the correct feature
matching result can be obtained after using the method of eliminating the mismatch above.
Then, the method of EPnP (Efficient Perspective-n-Point) is used to estimate the motion of the
camera between two adjacent frames, and then the estimated pose is optimized by the method
of graph optimization.
(3) The function of loop closing, global optimization and environment map construction of
the system is implemented. The concept of key frame is defined first. The key frame is selected
based on the bag of words to perform loop closing detection so that problem of position drift
caused by error accumulation can be solved. The global pose optimization is performed using
graph optimization after the loop closing, and then the globally consistent trajectories and maps
are constructed.
(4) In order to verify the validity and practicability of the system in this paper, experiments
were carried out using datasets and real environment. First, the experiments in the TUM dataset
万方数据
评论0