哈尔滨工业大学工学硕士学位论文
- II -
Abstract
SLAM(simultaneous localization and mapping) algorithm has an important
position in the robot science. In the field of robot navigation, it is necessary to acquire
the relative position relationship between the robot and the environment object in real
time, and then establish the coordinate system which includes the robot and the
environment object. This is the problem that the SLAM algorithm needs to solve. In
recent years, the AR / VR field, which is continuously heating up, also needs the SLAM
algorithm to compute the position of the user, and makes the SLAM problem become a
hotspot.
In the real world, the sensors of the robot always have errors, which leads to the
difficulty of obtaining a high-precision location information. It is usually necessary to
fuse, filter and optimize the sensor data. And The relationship between the sensor data
and the motion of the robot is usually not obvious; the amount of data that needs to be
optimized grows linearly with time; finally, In the long-term use, the error will continue
to accumulate, need to detect the existence of the trajectory loop to eliminate the
cumulative error.
For a long time, many scholars have done a lot of attempts on the SLAM problem,
from the 1990s SLAM concept was formally proposed to the current SLAM algorithm
flourishing; from the initial laser, sonar sensors, to the visual sensor, and then To a
variety of sensor fusion; from the use of filtering methods to achieve state estimation, to
the present use of optimization to achieve full SLAM optimization, SLAM problem
research has made great progress.
In this paper, the monocular vision SLAM algorithm based on contour features is
studied. In the current visual SLAM algorithm, the feature points are used as the visual
feature, but the point feature has some shortcomings, such as the need for the
complexity of the color change around the feature point, a large amount of calculation,
and so on. While using the contour as a feature can alleviate these problems to a certain
extent. Corresponding with the above problem, the number of contours in the image is
much less than the number of feature points. When the contour scale changes, many
features remain unchanged. Generally, the image pyramid is not needed. So the contour
extraction is much smaller than the feature extraction in the complexity of calculation.
Contour feature does not require the color complexity of the local area, as long as there
is gradient to extract the contour.
Keywords: SLAM, contour feature, computer vision
万方数据
评论0