Lecture 7: Localization

Question 1

What is Map-based localization?

Answer 1

The robot estimates its position using perceived information and a map

Question 2

What about the 2 things of the map used in map-based localization?

Answer 2

It might be known.
It might be built in parallel(simultaneous localization and mapping-SLAM)

Question 3

What are the 2 challenges of localization?

Answer 3

1)Measurements and the map are inherently error prone
2)The robot has to deal with uncertain information

Question 4

What is the approach of Localization?

Answer 4

The robot estimates the belief state about its position through an SEE and ACT cycle

Question 5

Measurements are error prone because of the following 3 things:

Answer 5

Odometry
Exteroceptive sensors(camera, laser)
Map

Question 6

The first See involves what?

Answer 6

The robot queries its sensors -> finds itself next to a pillar

Question 7

The Act involves what?

Answer 7

Robot moves one meter forward
* Motion estimated by wheel encoders
* Accumulation of uncertainty

Question 8

The second See involves what?

Answer 8

The robot queries its sensors again -> finds itself next to a pillar

Question 9

What does the belief update involve?

Answer 9

Information Fusion

Question 10

What are the 4 types of maps?

Answer 10

-Continuous map with single hypothesis
probability distribution 𝑝(𝑥)
-Continuous map with multiple hypotheses
probability distribution 𝑝(𝑥)
-Discretized metric map (grid 𝑘) with
probability distribution 𝑝(𝑘)
-Discretized topological map (nodes 𝑛) with
probability distribution 𝑝(𝑛).

Question 11

What does the Bayes rule relate and how is it written?

Answer 11

the conditional probability p(x|y) to its inverse p(y|x).
p(x|y) = np(y|x)p(x); n = p(y)^-1 normalization factor(integral of p = 1)

Question 12

Who uses the Bayes Rule and where is the Bayes rule theorem used?

Answer 12

The theorem is used by both Markov and Kalman-filter localization algorithms during the measurement update.

Question 13

Concerning application of theorem of total probability/convolution, what two probabilities is it applied to? What is the difference?

Answer 13

Continuous and Discrete probabilities. Continuous probabilities has dx(subscript)t-1

Question 14

What is Markov localization for?

Answer 14

Discretized pose representation

Question 15

What is Kalman filter for?

Answer 15

Continuous pose representation and Gaussian error model

Question 16

What is the SLAM problem?

Answer 16

how can a body navigate in a previously unknown environment, while constantly building and updating a map of its workspace using onboard sensors & onboard computation

Question 17

In what 4 situations is SLAM necessary?

Answer 17

-When a robot must be truly autonomous (no human input)
* When there is no prior knowledge about the environment
* When we cannot rely exclusively on pre-placed beacons or external
positioning systems (e.g., GPS)
* When the robot needs to know where it is

Question 18

How do you track the motion of a robot while it is moving?

Answer 18

SLAM(Simultaneous Localization and Mapping)

Question 19

SLAM is the backbone of what?

Answer 19

the spatial awareness of a robot

Question 20

The spatial awareness of a robot is what?

Answer 20

one of the most challenging problems in probabilistic robotics.

Question 21

An unbiased map is necessary for what?

Answer 21

Localizing the robot; pure localization with a known map

Question 22

An accurate pose estimates is necessary for?

Answer 22

building a map of the environment; Mapping with known robot poses.

Question 23

What makes SLAM better than an unbiased map and an accurate pose estimate?

Answer 23

SLAM has no prior knowledge of the robot’s workspace. The robot poses have to be estimated along the way.

Question 24

Where does SLAM originate from?

Answer 24

efforts to formalize production of topographic maps from aerial imagery

Question 25

What is “photogrammetry”?

Answer 25

The practice of determining the geometric properties of objects from images

Question 26

What is the goal of Photogrammetry?

Answer 26

align the images to build a topographic map of the area

Question 27

How does traditional SLAM work?

Answer 27

Pick natural scene features as landmarks, observe their motion & reason about robot motion.

Question 28

What does traditional SLAM research into?

Answer 28

Good features to track, sensors, trackers, representations, assumptions
Ways of dealing with uncertainty in the processes involved

Question 29

What are the 3 approaches to SLAM?

Answer 29

Bundle adjustment
Filtering(UKF/EKF/Particle Filter SLAM)
Keyframes

Question 30

What is the Vision from SLAM?

Answer 30

Images = information-rich snapshots of a scene
Compactness + affordability of cameras
HW advances

Question 31

When SLAM using a single,handheld camera?

Answer 31

Very applicable, compact, affordable

Question 32

What about Structure from Motion(SFM)?

Answer 32

Take some images of the object/scene to reconstruct
* Features (points, lines, …) are extracted from all frames and matched among
them.
* Process all images simultaneously: images do not need to be ordered.
* Optimization to recover both:
* Camera motion and
* 3D structure
* Up to a scale factor
* Not real-time

Question 33

MonoSLAM | problem statement

Answer 33

How can we track the motion of a camera while it is moving? i.e., online
Extract Shi-Tomasi features & track them in image space.
SLAM using a single camera, grabbing frames at 30Hz
Ellipses (in camera view) and Ellipsoids (in map view) represent uncertainty

Question 34

What about ORB-SLAM

Answer 34

The most powerful open-source
monocular SLAM approach today.
Uses ORB features (binary) in a
keyframe-based approach.
Binary place recognition

Question 35

What about OKVIS: open keyframe-based Visual Inertial SLAM

Answer 35

Visual-inertial SLAM odometry approach (i.e.
no loop-closure)
* Uses BRISK features in a keyframe-based
approach
* Tight visual-inertial fusion – handles both
monocular and stereo vision

Question 36

What about ROVIO?

Answer 36

• EKF-based
• Detects a variant of Shi-Tomasi
  features at different scale levels
• Tracks patches and uses the intensity
  errors in the innovation term
• Can only track a limited no. features,
  so ROVIO performs odometry, not
  SLAM.

Question 37

What about VINS-mono?

Answer 37

• a robust and versatile monocular visual-inertial state estimator
• Keyframe-based approach
• Runs SLAM based on a tightly- coupled visual-inertial odometry with relocalization.
• Shi-Tomasi features tracked using the KLT sparse optical flow tracker
• BRIEF descriptors for relocalization, using binary place recognition
• Extensions to stereo and IMU open- sourced.

Question 38

What’s next for SLAM

Answer 38

• Centralized: all data needs to pass through the server
  Remove redundant data
  Work towards distributed collaboration
• Server accessibility limits mission range
  Use cloud computing or a mobile server
  Or peer-to-peer communication in a distributed architecture.
• Enable stronger collaboration