哈尔滨工业大学工学硕士学位论文
Abstract
Simultaneous Localization and Mapping (SLAM) is a 3D reconstruction method. It
is a method of self-localization and mapping in an unknown environment. SLAM has a
wide range of applications in areas such as automatic driving and virtual reality.
Among various SLAM algorithms, the LSD-SLAM algorithm is a method that can
reconstruct large-scale scenes in real time. However, there is room for improvement in
the LSD-SLAM algorithm. In recent years, learning-based methods have been active in
various fields and have achieved good results. Therefore, this project attempts to use
learning-based methods to improve the robustness of LSD-SLAM.
The SLAM system consists of a sensor, a front-stage visual odometry, back-stage
nonlinear optimization, loop closure, and mapping. Among them, the visual odometry
part and the loop closure part can join the learning-based method. This subject has im-
proved the LSD-SLAM algorithm from these two aspects. Firstly, this topic proposes a
depth confidence estimation algorithm based on ground control points, which improves
the visual odometry part of the LSD-SLAM. The confidence estimation algorithm uses a
random forest algorithm to train the gorund control points prediction model. The mod-
el can use the features that can be calculated conveniently and quickly in the process of
stereo matching, so that the model costs as little as possible while prediction. In the pro-
cess of tracking estimation, an estimate of the depth confidence is obtained through the
model, and the confidence is used as a weight into the depth estimation and the camer-
a motion estimation to obtain a more accurate estimation result and improve the LSD-
SLAM algorithm in the front-stage visual odometry part accuracy, and then improve the
reconstruction accuracy of the entire system.
This topic also proposes a loop closure detection network model based on second-
order features. The model uses a deep convolutional network, based on the second-order
information, to obtain high-precision loop closure detection results. The loss function
used by the model is the triplet loss function. Through this kind of weakly supervised
learning, the distance between the features of the same place is continuously reduced, and
the feature distances of different locations are constantly increased, so that the features of
the same place keep clustering. The proposed loop closure detection network model is put
- II -
评论0