P1
Good afternoon dear teachers and classmates. My name is Jiayao Ma. It is my great honor to be here to present my work about ” SLAM for Indoor Mobile Robot with Contour Tracking
and Planar mapping” Thanks to my dear Prof. Hong.Liu . And thank you all to attend my master thesis defense.
Then I will introduce my work in following aspects:
First, the introduction
P2-p3
robotic is in its spring time. the picture shows the boston dynamic robot is searching in the forest.
Today our main topic is SLAM problem, it is an important part to let robot has the ability to navigate.
• Considering the robot see and act cycle,
the input of slam is the sensor data from the real world.
the output of slam is the map and robot localization.

Overall, SLAM shorts for Simultaneously Location and Mapping, it simultaneous localization and incremental mapping of the unknown environment from the unknown pose, while the robot moves.

It answer the question of where am i and what around me.

SLAM has wide applications in the fields of autonomous driving, indoor service robot,
not only in different kinds of robot,but also virtual or augmented reality. Which need the 3D estimation of the surrounding scence and the pose of the device form sensing data, sequentially and in real-time.

P5
one scope to classify slam system is by the kinds of sensors SLAM use.
Here, the figure shows the research trend and price of different sensors, we can see that,
The indoor robot using laser is the most mature solution, while the limit of laser is the high price and the small scope of perception. Nearly, the visual slam become a research trend.
In our work, we pay attention to the monucular camera based SLAM systemWhile limited by the high computation of image information.
in our research we focus on the indoor ground robot which only use the monocular camera.

benefits from its low sensor cost, researchers have conducted extensive research on it.

P6
while the monocular SLAM still has lots of difficulties for us to deal with, it is still a open problem, it. in location, map and other fields list in the bottom.
for example, two important difficulty , how to track camera in scene lack of texture and how to dense map in real-time.

this two question is under our attention.

P7
Hotspots:
most new releases of Visual SLAM systems are non-filter-based
In the non filter system, it can be considered as direct slam and non-direct slam by what information it use to match query image.
each method has its disadvantages,
The indirect method depends on scene texture , while the direct method performs better in texture-lack sence but still limits with robustness.
unfortunaly, The map of the two method are both sparse or semi-dense.
There are method could build dense map, while the computation are often high, usually depend on hardware GPU acceleration.

In this part we introudce the SLAM, defination, application, difficulties.
Now let's get into the second part, our research approach.

P8
before introduce our work,
a typical slam pipeline shown in berief.
there are 3 components.
visual odometer estimate camera motion from image sequences.
map reconstruction components build a consistent map with the camera pose and the image sequences.
the loop closing and back end component consistently optimize the whole pipeline.

P9

we motivated by he need of a dense map in real-time for robot to navaget and the computation resource limitaion on mobile robot platform.

the indoor environment means it would contain many man-made things in the environment
not like the laser, camera is more cheap and useful for the robot,
as well as we should concern the computation of SLAM.

we provide a slam algorithm that fuse planar contour with camera tracking as well as in the add a parallel thread fusing planar with sparse map to get dense reconstruction.

P10
firstly let's see how we do camera pose estimation with contour tracking.

Tradition method use the raw pixel intensity values to estimate a map of the environment and the camera motion. Instead , we use planar contour.
Firstly traditional method are based on the brightness consistency constraint and constant velocity motion model.
the first model said that
the projection of the 3D point P on the images should have consistent brightness .
the second model assumes a smooth camera motion.

Basically Two-view geometry, then Minimization the gray scale values of pixels. Assume the camera moves slowly, smoothly and the light condition does not change much.

we take the use of semantic man-made prior for low-gradient areas to complete map.
• Now let me introduce our research approach.
What is super pixel, Superpixels are middle-level features consisting of image regions of homogeneous texture.
We assume that superpixels correspond to planar surfaces, use a homography model and minimize the distance between the contours.
The traditional method using superpixel can’t be solve in realtime for the reason that
1 their repeatability is low and highly dependent on specularities, camera bright/gain and the content of the image.
2 due to this low repeatability
and the lack of robust descriptors, the matching is difficult.

P13
two important issue with camera pose estimation is the pinhole model and two view geometry,
the first model said the relationship between the point in the image and the real world under camera view.
the second model said the relationship between the camera motion with the projection of 3d point on the frames.

P14
the camera pose estimation based on two model ,
the first model said that the projection of the 3d point P on the image should have the same brightness.
the second model assumes a smooth camera motion.
Based on pinhole model and two view geometry illustrated before,
The photo metric error function on the point pu with image i,
are related with
F is from the pinhole camera model
T is the translation between cuttrent frame , last key frame and global frame.

P15

Where J F r are the gradients of the residual reference
keyframe, J T F k is the derivative of the projection model with
T
respect to the transformationT k and J ε k is the derivative of
the transformation T k with respect to the motion ε.

P16
P17
we consider every pixel in the image is a vertex in the graph

P20

First to measure the similarity between patches.
Calculate all the patches on the eipolar line to get the distribution on the polar lines

Then , filter based depth estimate the distribution is a Gauss distribution
the depth d of pixel

Then , filter based depth estimate the distribution is a Gauss distribution
the depth d of pixel
Each observation
Depth Observation Fusion N(μfuse,σ_fuse^2 ) Should be
minimizing the photo metric error for several overlapping views.
minimizing the photo metric error