1. 16722 Sensing and Sensors Mel Siegel (mws@cmu.edu, 412-983-2626)
Sensing and Sensors 2004 Spring Lecture 1
16722 Sensing and Sensors
Mel Siegel )
/afs/cs/academic/16722/S2004
course description
intro to sensor data processing and sensor fusion

2. 16-722 Sensing and Sensors 2004 Spring Lecture 1
Class structure
09:30 AM – 10:10 AM administration stuff, background lecture
10:10 AM – 10:50 AM student lecture 1 (30’ lecture + 10’ discussion)
10:50 AM – 11:00 AM break
11:00 AM – 11:40 AM student lecture 2 (30’ lecture + 10’ discussion)
10:40 AM – 12:20PM student lecture 3 (30’ lecture + 10’ discussion)
Lecture slides, well annotated
1st content slide: bibliography
Last slide: bibliography for further reading
Distribute: 10 short quiz (50 pts) + 1 quantitative/programming (50 pts)
Collect answers: Monday noon, graded to Mel: Wednesday noon
Lecture: 100 (Mel) (peers) follow-up: 100 (Mel) (peers)
Final exams can be exempt if you will clearly get an “A”!
p.s. all rights reserved regarding its interpretations/changes

3. 16-722 Sensing and Sensors 2004 Spring Lecture 1
Sensing  Sensors
sensing: situation awareness
understanding it from its underlying physics concepts:
measurement limitation
signal v.s. noise
sensor: implementing sensing
sensor fusion: obtain

4. Sensory Data Preprocessing
Sensing and Sensors 2004 Spring Lecture 1
Sensory Data Preprocessing
to know your sensor’s behavior: measure the same parameter conceptually under unchanged condition:
accuracy: how far away is from the “true” value ?
precision: how big is the variation s ?
 calibrate sensor to eliminate system error
 using s estimation to kick out out-lines
 take multiple measurements to lower variation:

5. Introduction to Sensor Fusion
Sensing and Sensors 2004 Spring Lecture 1
Introduction to Sensor Fusion
multiple sensors are needed, how to obtain the best “truth” in terms of actual utility
U.S. Joint Directors of Laboratories (JDL) Data Fusion Group, 1985: a process dealing with the association, correlation, and combination of data and information from single and multiple sources to achieve a refined position and identity estimates, and complete and timely assessments of situations and threats, and their significance. The process is characterized by continuous refinements of its estimates and assessments, and the evaluation of the need for additional sources, or modification of the process itself, to achieve improved results.
the process of combining data or information to estimate or predict entity states

6. sub-object assessment database management system
Sensing and Sensors 2004 Spring Lecture 1
Sensor Fusion Model
the revised JDL data fusion model (1998)
Level 0
sub-object assessment
Level 1
object assessment
Level 2
situation assessment
Level 3
impact assessment
Human-
Computer
Interface
external
Sensors
Documents
People .
Data stores
distributed
Sensors
Documents
People .
Data stores
local
Sensors
Documents
People .
Data stores
Level 4
process refinement
database management system
support
database
fusion
database
sources

7. Sensor Fusion Architectures
Sensing and Sensors 2004 Spring Lecture 1
Sensor Fusion Architectures
sensors
joint identity declaration S1 Sn S2
association
Data Level Fusion
feature extraction
identity declaration
(a)
Feature Level Fusion identity declaration
(b)
id decl
Declaration Level Fusion identity declaration
(c)

8. Sensor Data Fusion Types
Sensing and Sensors 2004 Spring Lecture 1
Sensor Data Fusion Types
sensors’ output may be “competitive”, “compliment”, or “corporative”
(Forward-Looking Infrared image sensor)

9. “objective” Sensor Fusion Techniques
Sensing and Sensors 2004 Spring Lecture 1
“objective” Sensor Fusion Techniques
naive statistical inference:
voting fusion mechanism:

10. “subjective” Sensor Fusion: Bayesian inference framework
Sensing and Sensors 2004 Spring Lecture 1
“subjective” Sensor Fusion: Bayesian inference framework
basic assumption: hypotheses exclusive & exhaustive
different explanations of Bayesian framework

11. “subjective” Sensor Fusion: Dempster-Shafer framework
Sensing and Sensors 2004 Spring Lecture 1
“subjective” Sensor Fusion: Dempster-Shafer framework
frame of discernment:
reasoning space: power-set of the basic elements
basic belief assignment does not care complements:
 belief, plausibility, ambiguity, & ignorance are modeled

12. Dempster-Shafer Evidence Combination rule
Sensing and Sensors 2004 Spring Lecture 1
Dempster-Shafer Evidence Combination rule
{L}=0.3
{R}=0.6
{L|R}=0.1
{L}=0.4
{L}=0.4x0.3
{Φ}=0.4x0.6
{L}=0.4x0.1
{R}=0.5
{Φ}=0.5x0.3
{R}=0.5x0.6
{L}=0.1x0.3
{R}=0.1x0.6
{L|R}=0.1x0.1

13. Dempster-Shafer for Practical Usage
Sensing and Sensors 2004 Spring Lecture 1
Dempster-Shafer for Practical Usage
participants are rational & consistent: democratic voting
reality: some sensors are more reliable & should enjoy more credibility

14. Dempster-Shafer Extension
Sensing and Sensors 2004 Spring Lecture 1
Dempster-Shafer Extension
performance over time factor: drift to be considered

15. Bayesian versus Dempster-Shafer
Sensing and Sensors 2004 Spring Lecture 1
Bayesian versus Dempster-Shafer
Bayesian
Dempster-Shafer
hypothesis must be exclusive
some hypotheses can include others -- different granularity information can be included
prior knowledge & new observations unambiguously relate to probabilities
prior probability unknown, new observations partially relate to probability & ignorance be counted
joint probability distribution must be available
joint probability distribution is not required
direct correlation with probability helps maximizing expected utility
skills needed to correlate evidences with probability, thus weak in decision-making support
a generalization/extension of Bayesian inference network