哈尔滨工业大学工程硕士学位论文
II
Abstract
With the development of global technology, mobile robotics is not limited to
academic research and slowly penetrates into all fields, such as industry, medical
treatment, military, adventure, household products and so on. Intelligent robot can sense
and recognize the surrounding environment. Through the surrounding environment, the
robot can carry out self-learning, dynamic decision and planning. However, robot
navigation is the basic problem of the mobile robots, and the important research
direction of the mobile robots. If the robots want to achieve self-navigation,
self-positioning, mapping and path planning are the main problems that should be
solved firstly.
In my article, the self-positioning and mapping are solved through the RGB-D
SLAM frame, it includes image processing front and optimized backend. at the front
part, For the feature detection and extraction, the Kinect sensor often has appearance
noise and depth noise when it gets image data, the traditional visual-only features are
often affected, a novel approach is proposed in my article that first extracts planes from
the point cloud, acquired by the RGB-D camera, the appearance noise can be filtered,
then tries to enhance the texture on the extracted planes, the depth of features can be
corrected through the parameters of plane models, it can reduce the effect of depth noise,
then point features are then detected. For the feature matching, the traditional algorithm,
BF and FLANN has some wrong matches, so the KNN based RANSAC is used to
remove the wrong matching, the KNN can detect the duplex matching, using
homography transformation to optimize the matching results, and using RANSAC to
filter the rest part.
For the motion estimation, using ICP based RANSAC to estimate the parameter of
robot’s motion, because the traditional algorithm easily make the ICP fall into the local
minimum. The optimized backend includes building the pose map through the front data,
eliminating wrong loop closure and pose graph optimization using g2o.
For the robot path planning, using the 16 direction instead of 8 direction based A*
algorithm, could reduce the length of actual running route and the sum of rotation angle,
using the improved method could reduce the energy of robot, at last, using the improved
A* algorithm to perform global planning and using the dynamic A* algorithm to
perform local planning.
Keywords: RGB-D SLAM; path planning; feature detection; emotion estimation
万方数据
