Principles of Robot Autonomy I

Robotic sensors and introduction to computer vision

Sensors for mobile robots
Knowledge Mission goals
Localiza(on posi(on Decision making
Map Building global map Mo(on planning

environmental model trajectory

local map

Informa(on Trajectory
extrac(on execu(on

actuator
raw data commands
See-think-act
Sensing Actua(on

Real world
environment

10/12/21 AA 274 | Lecture 7 2

Sensors for mobile robots
• Aim
• Learn about key performance characteristics for robotic sensors
• Learn about a full spectrum of sensors, e.g. proprioceptive / exteroceptive,
passive / active

• Readings
• Siegwart, Nourbakhsh, Scaramuzza. Introduction to Autonomous Mobile
Robots. Section 4.1.

10/12/21 AA 274 | Lecture 7 3

Example: self-driving cars

10/12/21 AA 274 | Lecture 7 4

10/12/21 AA 274 | Lecture 7 5
Classification of sensors
• Proprioceptive: measure values internal to the robot
• E.g.: motor speed, robot arm joint angles, and battery voltage
• Exteroceptive: acquire information from the robot’s environment
• E.g.: distance measurements and light intensity

• Passive: measure ambient environmental energy entering the

sensor
• Challenge: performance heavily depends on the environment
• E.g.: temperature probes and cameras
• Active: emit energy into the environment and measure the reaction
• Challenge: might affect the environment
• E.g.: ultrasonic sensors and laser rangefinders
10/12/21 AA 274 | Lecture 7 6
Sensor performance: design specs
• Dynamic range: ratio between the maximum and minimum input
values (for normal sensor operation)

• Resolution: minimum difference between two values that can be

detected by a sensor

• Linearity: whether or not the sensor’s output response depends

linearly on the input

• Bandwidth or frequency: speed at which a sensor provides readings

(in Hertz)

10/12/21 AA 274 | Lecture 7 7

Sensor performance: in situ specs
• Sensitivity: ratio of output change to input change

• Cross-sensitivity: sensitivity to quantities that are unrelated to the

target quantity

• Error: difference between the sensor output m and the true value v
error ≔ 𝑚 − 𝑣

• Accuracy: degree of conformity between the sensor’s measurement

and the true value
accuracy ≔ 1 − |error|/𝑣

• Precision: reproducibility of the sensor results

10/12/21 AA 274 | Lecture 7 8
Sensor errors
• Systematic errors: caused by factors that can in theory be modeled;
they are deterministic
• E.g.: calibration errors

• Random errors: cannot be predicted with sophisticated models;

they are stochastic
• E.g.: spurious range-finding errors

• Error analysis: performed via a probabilistic analysis

• Common assumption: symmetric, unimodal (and often Gaussian)
distributions; convenient, but often a coarse simplification
• Error propagation characterized by the error propagation law
10/12/21 AA 274 | Lecture 7 9
An ecosystem of sensors
• Encoders
• Heading sensors
• Accelerometers and IMU
• Beacons
• Active ranging
• Cameras

10/12/21 AA 274 | Lecture 7 10

Encoders
• Encoder: an electro-mechanical device that Wheel encoder
converts motion into a sequence of digital Credit: Pololu

pulses, which can be converted to relative or

absolute position measurements
• proprioceptive sensor
• can be used for robot localization

• Fundamental principle of optical encoders: use

a light shining onto a photodiode through slits
in a metal or glass disc

10/12/21 AA 274 | Lecture 7 Credit: Honest Sensor 11

Heading sensors
• Used to determine robot’s orientation, it can be:
1. Proprioceptive, e.g., gyroscope (heading sensor that preserves its
orientation in relation to a fixed reference frame)
2. Exteroceptive, e.g., compass (shows direction relative to the geographic
cardinal directions)
• Fusing measurements with velocity information, one can obtain a
position estimate (via integration) -> dead reckoning

• Fundamental principle of mechanical

gyroscopes: angular momentum associated
with spinning wheel keeps the axis of
rotation inertially stable
10/12/21 AA 274 | Lecture 7 Credit: SNS 12
Accelerometer and IMU
• Accelerometer: device that measures all external forces acting upon it
• Mechanical accelerometer: essentially, a spring-mass-damper system

with m mass of proof mass, c damping coefficient,

k spring constant; in steady state

• Modern accelerometers use MEMS (cantilevered

beam + proof mass); deflection measured via
capacitive or piezoelectric effects
10/12/21 AA 274 | Lecture 7 13
Inertial Measurement Unit (IMU)
• Definition: device that uses gyroscopes and accelerometers to
estimate the relative position, orientation, velocity, and
acceleration of a moving vehicle with respect to an inertial frame
• Drift is a fundamental problem: to cancel drift, periodic references
to external measurements are required

Credit: SNS

10/12/21 AA 274 | Lecture 7 14

Beacons

• Definition: signaling devices

with precisely known positions
• Early examples: stars,
lighthouses
• Modern examples: GPS,
motion capture systems

10/12/21 AA 274 | Lecture 7 15

Active ranging
• Provide direct measurements of distance to objects in vicinity
• Key elements for both localization and environment reconstruction
• Main types:
1. Time-of-flight active ranging sensors (e.g., ultrasonic and laser rangefinder)

Credit:
https://electrosome.c
om/hc-sr04-
ultrasonic-sensor-pic/

2. Geometric active ranging sensors (optical triangulation and structured light)

10/12/21 AA 274 | Lecture 7 16

Time-of-flight active ranging
• Fundamental principle: time-of-flight ranging makes use of the
propagation of the speed of sound or of an electromagnetic wave
• Travel distance is given by

where d is the distance traveled, c is the speed of the wave

propagation, and t is the time of flight
• Propagation speeds:
• Sound: 0.3 m/ms
• Light: 0.3 m/ns
• Performance depends on several factors, e.g., uncertainties in
determining the exact time of arrival and interaction with the target
10/12/21 AA 274 | Lecture 7 17
Geometric active ranging
• Fundamental principle: use geometric properties in the
measurements to establish distance readings
• The sensor projects a known light pattern (e.g., point, line, or
texture); the reflection is captured by a receiver and, together with
known geometric values, range is estimated via triangulation
• Examples:
• Optical triangulation (1D sensor)
• Structured light (2D and 3D sensor)

Credit: Matt Fisher

10/12/21 AA 274 | Lecture 7 18

Several other sensors are available
• Classical, e.g.:
• Radar (possibly using Doppler effect to produce velocity data)
• Tactile sensors

• Emerging technologies:
• Artificial skins
• Neuromorphic cameras

10/12/21 AA 274 | Lecture 7 19

Introduction to computer vision
• Aim
• Learn about cameras and camera models

• Readings
• Siegwart, Nourbakhsh, Scaramuzza. Introduction to Autonomous Mobile
Robots. Section 4.2.3.
• D. A. Forsyth and J. Ponce [FP]. Computer Vision: A Modern Approach (2nd
Edition). Prentice Hall, 2011. Chapter 1.
• R. Hartley and A. Zisserman [HZ]. Multiple View Geometry in Computer
Vision. Academic Press, 2002. Chapter 6.1.
10/12/21 AA 274 | Lecture 7 20
Vision
• Vision: ability to interpret the surrounding environment using light in
the visible spectrum reflected by objects in the environment
• Human eye: provides enormous amount of information, ~millions of
bits per second
• Cameras (e.g., CCD, CMOS): capture light -> convert to digital image
-> process to get relevant information (from geometric to semantic)
Sky

1. Information extraction
2. Interpretation

Car
10/12/21 AA 274 | Lecture 7 21
How to capture an image of the world?
• Light is reflected by the object and scattered in all directions
• If we simply add a photoreceptive surface, the captured image will be
extremely blurred

Photoreceptive surface

10/12/21 AA 274 | Lecture 7 22

Pinhole camera
• Idea: add a barrier to block off most of the rays

Photoreceptive surface

Aperture

• Pinhole camera: a camera without a lens but with a tiny aperture, a

pinhole
10/12/21 AA 274 | Lecture 7 23
A long history
• Very old idea (several thousands of years BC)
• First clear description from Leonardo Da Vinci (1502)
• Oldest known published drawing of a camera obscura by Gemma
Frisius (1544)

10/12/21 AA 274 | Lecture 7 24

Pinhole camera
Focal length
Pinhole of aperture

Credit: FP Chapter 1

• Perspective projection creates inverted images

• Sometimes it is convenient to consider a virtual image associated
with a plane lying in front of the pinhole
• Virtual image not inverted but otherwise equivalent to the actual one
10/12/21 AA 274 | Lecture 7 25
Pinhole perspective
Image plane
Pinhole
Image center

Perspective

Credit: FP Chapter 1
Optical axis
• Since P, O, and p are collinear: 𝑂𝑝 = 𝜆𝑂𝑃 for some 𝜆 ∈ 𝑅
• Also, z=f, hence

10/12/21 AA 274 | Lecture 7 26

Issues with pinhole camera
• Larger aperture -> greater
number of light rays that pass
through the aperture -> blur
• Smaller aperture -> fewer
number of light rays that pass
through the aperture ->
darkness (+ diffraction)

• Solution: add a lens to replace

the aperture!

10/12/21 AA 274 | Lecture 7 27

Lenses
• Lens: an optical element that focuses light by means of refraction

10/12/21 AA 274 | Lecture 7 28

Thin lens model
Key properties (follows from Snell’s law) :
1. Rays passing through O are not
refracted
2. Rays parallel to the optical axis are
focused on the focal point Fʹ
3. All rays passing through P are focused
by the thin lens on the point p

Credit: FP Chapter 1

• Similar triangles
Blue triangles
Thin lens
equation
Red triangles

10/12/21 AA 274 | Lecture 7 29

Thin lens model
• Key points:
1. The equations relating the positions of P and p are exactly the same as
under pinhole perspective if one considers z as focal length (as opposed
to f), since P and p lie on a ray passing through the center of the lens
2. Points located at a distance −Z from O will be in sharp focus only when
the image plane is located at a distance z from O on the other side of the
lens that satisfies the thin lens equation
3. In practice, objects within some range of distances (called depth of field
or depth of focus) will be in acceptable focus
4. Letting 𝑍 → ∞ shows that f is the distance between the center of the lens
and the plane where distant objects focus
5. In reality, lenses suffer from a number of aberrations

10/12/21 AA 274 | Lecture 7 30

 Perspective projection
• Goal: find how world points map in the camera image
• Assumption: pinhole camera model (all results also hold under thin
lens model, assuming camera is focused at ∞)
P in camera frame
Camera frame Roadmap:
P in world frame 1. Map 𝑃! into p (image plane)
2. Map p into (u,v) (pixel coordinates)
3. Transform 𝑃" into 𝑃!

Image plane World frame

10/12/21 AA 274 | Lecture 7 31
Step 1
• Task: Map 𝑃! = (𝑋" , 𝑌" , 𝑍" ) into 𝑝 = (𝑥, 𝑦) (image plane)
• From before

10/12/21 AA 274 | Lecture 7 32

Step 2.a
• Actual origin of the camera coordinate system is usually at a corner
(e.g., top left, bottom left)

10/12/21 AA 274 | Lecture 7 33

Step 2.b
• Task: convert from image coordinates (𝑥, ? to pixel coordinates
? 𝑦)
(𝑢, 𝑣)
• Let 𝑘# and 𝑘$ be the number of pixels per unit distance in image
coordinates in the x and y directions, respectively

Nonlinear transformation
10/12/21 AA 274 | Lecture 7 34
 Homogeneous coordinates
• Goal: represent the transformation as a linear mapping
• Key idea: introduce homogeneous coordinates

Inhomogenous -> homogeneous Homogenous -> inhomogeneous

10/12/21 AA 274 | Lecture 7 35

Perspective projection in
homogeneous coordinates
• Projection can be equivalently written in homogeneous coordinates

Camera matrix/ 𝑃! in homogeneous Homogeneous pixel

Matrix of intrinsic parameters coordinates coordinates

• In homogeneous coordinates, the mapping is linear:

Point p in homogeneous Point 𝑃! in homogeneous
pixel coordinates camera coordinates
10/12/21 AA 274 | Lecture 7 36
Skewness
• In some (rare) cases
Skew parameter

• When is 𝛾 ≠ 0?
• x- and y-axis of the camera are not perpendicular (unlikely)
• For example, as a result of taking an image of an image
• Five parameters in total!

10/12/21 AA 274 | Lecture 7 37

Next time: camera models & calibration

10/12/21 AA 274 | Lecture 7 38