a networking perspective part 2 – option b · atx motherboard ! 1,6 ghz intel pentium 4, 256 mb...
TRANSCRIPT
![Page 1: a networking perspective Part 2 – option B · ATX motherboard ! 1,6 GHz Intel Pentium 4, 256 MB RAM, 160 GB HD ... Smart Object Identification ! The AMR gets a list of close by](https://reader036.vdocuments.us/reader036/viewer/2022071018/5fd1af70cd79d24a5660db8e/html5/thumbnails/1.jpg)
Signal processing: a networking perspective Part 2 – option B
Andrea Zanella
PhD Summer School of Information Engineering – 7-11 July, 2014 – Brixen (BZ)
[email protected] SIGnals processing & NETworking research group
1
![Page 2: a networking perspective Part 2 – option B · ATX motherboard ! 1,6 GHz Intel Pentium 4, 256 MB RAM, 160 GB HD ... Smart Object Identification ! The AMR gets a list of close by](https://reader036.vdocuments.us/reader036/viewer/2022071018/5fd1af70cd79d24a5660db8e/html5/thumbnails/2.jpg)
Outline
¨ Part 3: QoE-oriented networking ¤ Mixing up PHY and application-layer signals
n smart environments
¤ Context awareness from signal processing n SSIM characterization of video content n Handover optimization
¤ Bye bye!
http://www.keepcalm-o-matic.co.uk/p/keep-calm-it-s-lunch-time/
2
![Page 3: a networking perspective Part 2 – option B · ATX motherboard ! 1,6 GHz Intel Pentium 4, 256 MB RAM, 160 GB HD ... Smart Object Identification ! The AMR gets a list of close by](https://reader036.vdocuments.us/reader036/viewer/2022071018/5fd1af70cd79d24a5660db8e/html5/thumbnails/3.jpg)
Exploiting complementarities
¨ Pros ¤ Low cost
n Large networks with many devices ¤ Low energy consumption
n Possibility to power with small batteries or energy-scavenging
¨ Cons ¤ Limited computing capabilities ¤ Limited memory ¤ Limited transmission range/speed
¨ Cons ¤ High cost
n Single or few units ¤ High energy consumption
n Need for large batteries and/or recharging stations
¨ Pros ¤ High computational capabilities ¤ Large memory ¤ Large energy capacity ¤ Large transmission range/speed ¤ Advanced mobility
Wireless Sensor Nodes (WSN) Autonomous Mobile Robots (AMR)
3
![Page 4: a networking perspective Part 2 – option B · ATX motherboard ! 1,6 GHz Intel Pentium 4, 256 MB RAM, 160 GB HD ... Smart Object Identification ! The AMR gets a list of close by](https://reader036.vdocuments.us/reader036/viewer/2022071018/5fd1af70cd79d24a5660db8e/html5/thumbnails/4.jpg)
Application scenarios
¨ Smart Environments ¤ AMRs may catalog, localize and
retrieve objects upon request ¤ Libraries: AMRs may put back books on
the shelves after closure ¨ Assisted Living
¤ A portable device may drive a visual-impaired person through a partially unknown environment by vocally describing the smart objects that recognizes along the way
¤ A personal robot may display to a motion-impaired user the list of smart objects that it discovers n remote controls, mobile phones, books, …
¨ Dataset of object-models stored on the robot ¤ Low flexibility: new objects shall
be included in the directory to be recognized
¤ Similar objects are indistinguishable
¨ RFID-based communication ¤ Cheaper, but requires physical
proximity and does not support sophisticated algorithms nor multi-hop communication
Application Scenarios Limit of existing solutions
4
![Page 5: a networking perspective Part 2 – option B · ATX motherboard ! 1,6 GHz Intel Pentium 4, 256 MB RAM, 160 GB HD ... Smart Object Identification ! The AMR gets a list of close by](https://reader036.vdocuments.us/reader036/viewer/2022071018/5fd1af70cd79d24a5660db8e/html5/thumbnails/5.jpg)
Target application: autonomous exploration of unknown “smart” environments
¨ AMR equipped with ¤ odometers for self-localization
¤ on-board camera
¤ RF transceiver
¨ Smart Objects (SOs), i.e., common objects tagged with WSNs that provide ¤ Computational/communication capabilities
¤ data storage
¤ self-management features
¨ Autonomous exploration of unknown environments and interaction with SOs ¤ No prior info about the number or kind of SOs
Vision Goal
5
![Page 6: a networking perspective Part 2 – option B · ATX motherboard ! 1,6 GHz Intel Pentium 4, 256 MB RAM, 160 GB HD ... Smart Object Identification ! The AMR gets a list of close by](https://reader036.vdocuments.us/reader036/viewer/2022071018/5fd1af70cd79d24a5660db8e/html5/thumbnails/6.jpg)
Experimental Set up
¨ 250 kbps @ 2.4 GHz ¨ Ultra low current consumption
¨ Light, temperature, RSSI sensors
¨ TinyOS
¨ Home-made, based on Pioneer 2 ActivMedia platform
¨ Linux OS with Miro middleware ¨ ATX motherboard
¤ 1,6 GHz Intel Pentium 4, 256 MB RAM, 160 GB HD
¨ On-board camera ¨ Odometers
TmoteSky sensor nodes AMR Bender
TmoteSky connected to ATX via USB + MoteService class
added to Miro
6
![Page 7: a networking perspective Part 2 – option B · ATX motherboard ! 1,6 GHz Intel Pentium 4, 256 MB RAM, 160 GB HD ... Smart Object Identification ! The AMR gets a list of close by](https://reader036.vdocuments.us/reader036/viewer/2022071018/5fd1af70cd79d24a5660db8e/html5/thumbnails/7.jpg)
System’s building blocks
1. Smart Object Identification ¤ The AMR gets a list of close by SOs, but does not
know their location 2. Smart Object Mapping ¤ The AMR builds a map of the area estimating SOs’
location 3. Smart Object Recognition ¤ The AMR gets close to a selected object and, then,
visually recognizes it in the image taken by the on-board camera
7
![Page 8: a networking perspective Part 2 – option B · ATX motherboard ! 1,6 GHz Intel Pentium 4, 256 MB RAM, 160 GB HD ... Smart Object Identification ! The AMR gets a list of close by](https://reader036.vdocuments.us/reader036/viewer/2022071018/5fd1af70cd79d24a5660db8e/html5/thumbnails/8.jpg)
Step 1. Smart Objects Identification
8
![Page 9: a networking perspective Part 2 – option B · ATX motherboard ! 1,6 GHz Intel Pentium 4, 256 MB RAM, 160 GB HD ... Smart Object Identification ! The AMR gets a list of close by](https://reader036.vdocuments.us/reader036/viewer/2022071018/5fd1af70cd79d24a5660db8e/html5/thumbnails/9.jpg)
Identification: centralized approach
ID=1, Laptop
ID=2, cellphone
ID=3, ashtray
ID=4, keys
ID=5, Tablet PC
Who’s out there?
ID Object type
3 ashtray
5 Tablet PC
4 Keys
1 Laptop
2 Cell phone
Smart Objects Lists
9
![Page 10: a networking perspective Part 2 – option B · ATX motherboard ! 1,6 GHz Intel Pentium 4, 256 MB RAM, 160 GB HD ... Smart Object Identification ! The AMR gets a list of close by](https://reader036.vdocuments.us/reader036/viewer/2022071018/5fd1af70cd79d24a5660db8e/html5/thumbnails/10.jpg)
Identification: distributed approach
Hello! I’m ID=1, Laptop I know:
ID=3, ashtray ID=5, Tablet PC
ID=4, Keys
Hello, I’m ID=2, cellphone I know:
ID=3, ashtray ID=5, Tablet PC
ID=4, Keys ID=5, Laptop
Hello! I am: ID=3, ashtray
I know: nobody
Hello! I’m ID=4, keys I know:
ID=3, ashtray ID=5, Tablet PC
Hello! I’m: ID=5, Tablet PC I know: ID=3, ashtray
Who’s out there?
ID Object type
3 ashtray
5 Tablet PC
4 Keys
1 Laptop
2 Cell phone
Smart Objects Lists
Hello, I’m ID=2, cellphone I know:
ID=3, ashtray ID=5, Tablet PC
ID=4, Keys ID=5, Laptop
10
![Page 11: a networking perspective Part 2 – option B · ATX motherboard ! 1,6 GHz Intel Pentium 4, 256 MB RAM, 160 GB HD ... Smart Object Identification ! The AMR gets a list of close by](https://reader036.vdocuments.us/reader036/viewer/2022071018/5fd1af70cd79d24a5660db8e/html5/thumbnails/11.jpg)
Step 2. Smart Objects Mapping
11
![Page 12: a networking perspective Part 2 – option B · ATX motherboard ! 1,6 GHz Intel Pentium 4, 256 MB RAM, 160 GB HD ... Smart Object Identification ! The AMR gets a list of close by](https://reader036.vdocuments.us/reader036/viewer/2022071018/5fd1af70cd79d24a5660db8e/html5/thumbnails/12.jpg)
Mapping: still a challenge
¨ GPS works pretty well… outdoor… ¤ Not that well indoor though! ¤ Hungry for energy (& money): not suitable for SOs
¨ Cricket-like systems work fine, but require dedicated hw ¤ Impact on device size, cost, location, energy efficiency…
¨ RF-based localization techniques are very attractive… ¤ if you’re not too picky about performance!
n Radio Signal Strength Indication (RSSI) is natively provided by all WSN platforms
n However, RSSI-based ranging is usually VERY noisy!
12
![Page 13: a networking perspective Part 2 – option B · ATX motherboard ! 1,6 GHz Intel Pentium 4, 256 MB RAM, 160 GB HD ... Smart Object Identification ! The AMR gets a list of close by](https://reader036.vdocuments.us/reader036/viewer/2022071018/5fd1af70cd79d24a5660db8e/html5/thumbnails/13.jpg)
Mapping: our solution
¨ Improve RSSI-based ranging by using the communication capabilities of WSNs ¤ Multi-channel RSSI
averaging
¨ Use on-board odometers of the AMR and its computational capabilities to perform sophisticated localization algorithm ¤ SLAM: Simultaneous
Localization (of the AMR) And Mapping (of the SOs)
WSN-wise AMR-wise
13
![Page 14: a networking perspective Part 2 – option B · ATX motherboard ! 1,6 GHz Intel Pentium 4, 256 MB RAM, 160 GB HD ... Smart Object Identification ! The AMR gets a list of close by](https://reader036.vdocuments.us/reader036/viewer/2022071018/5fd1af70cd79d24a5660db8e/html5/thumbnails/14.jpg)
Received power
Transmit power Path loss coefficient
reference distance
environmental constant
transmitter-receiver distance
Deterministic component
RSSI-based ranging: basics
¨ Path loss channel model ¤ received power Pi @ distance di
Shadowing
Random term
Fast Fading
€
ˆ d i =10PTx +K−Pi [dBm ]
10η = di10−ψ i
10ηRanging error
Propagation-based ranging
14
![Page 15: a networking perspective Part 2 – option B · ATX motherboard ! 1,6 GHz Intel Pentium 4, 256 MB RAM, 160 GB HD ... Smart Object Identification ! The AMR gets a list of close by](https://reader036.vdocuments.us/reader036/viewer/2022071018/5fd1af70cd79d24a5660db8e/html5/thumbnails/15.jpg)
RSSI-ranging: an example…
Random variations due to shadowing and fading obscure the log-decreasing law for the received power vs distance
Distance [mt]
Rec
eive
d po
wer
Pi [
dBm
]
Measured samples Deterministic component
15
![Page 16: a networking perspective Part 2 – option B · ATX motherboard ! 1,6 GHz Intel Pentium 4, 256 MB RAM, 160 GB HD ... Smart Object Identification ! The AMR gets a list of close by](https://reader036.vdocuments.us/reader036/viewer/2022071018/5fd1af70cd79d24a5660db8e/html5/thumbnails/16.jpg)
Mitigating RSSI randomness
¨ Shadowing term is actually due to strong and slowly time varying multipath interference
¨ A shift in the carrier frequency will change the way reflected copies combine at the receiver
16
![Page 17: a networking perspective Part 2 – option B · ATX motherboard ! 1,6 GHz Intel Pentium 4, 256 MB RAM, 160 GB HD ... Smart Object Identification ! The AMR gets a list of close by](https://reader036.vdocuments.us/reader036/viewer/2022071018/5fd1af70cd79d24a5660db8e/html5/thumbnails/17.jpg)
Multi-channel RSSI: experimental results
Time Average Frequency average
17
! " # $ %!&!
!%!
!'!
!$!
!(!
!#!
!)!
!"!
*+,-./012345
6,,+237845
2
2
-+41!.91:.;172:,,2,.4<=1,
<.->!=?,24?71=
! " # $ %!$&
!$!
!&&
!&!
!#&
!#!
!'&
!'!
!"&
()*+,-./0123
4**)015623
0
0
78/9!,:/8,;/508**0*,2<=/*
<,+>!=?*02?5/=
![Page 18: a networking perspective Part 2 – option B · ATX motherboard ! 1,6 GHz Intel Pentium 4, 256 MB RAM, 160 GB HD ... Smart Object Identification ! The AMR gets a list of close by](https://reader036.vdocuments.us/reader036/viewer/2022071018/5fd1af70cd79d24a5660db8e/html5/thumbnails/18.jpg)
RF-jumping protocol in a nutshell!
Anybody in CH1? (Next CH2)
I’m in CH1! (Next CH2)
I’m in CH1! (Next CH2)
Everybody’s switching on CH2. Let’s follow them!
18
![Page 19: a networking perspective Part 2 – option B · ATX motherboard ! 1,6 GHz Intel Pentium 4, 256 MB RAM, 160 GB HD ... Smart Object Identification ! The AMR gets a list of close by](https://reader036.vdocuments.us/reader036/viewer/2022071018/5fd1af70cd79d24a5660db8e/html5/thumbnails/19.jpg)
RF-jumping protocol
¨ Set TX Channel to default RF Channel
¨ FOR NextRFCh=1 TO 16 DO ¤ Send REQUEST(NextRFCh)
n Start REQ_TO ¤ WHILE NOT(REQ_TO expired) DO
n IF (REPLY pkt received) n Store RSSI sample n Restart REQ TimeOut
¤ ENDWHILE ¤ Set TX Channel to NextRFCh
¨ ENDFOR
¨ Set flag=false ¨ IF (REQUEST pkt received)
¤ Get NextRFCh ¤ Set flag=true ¤ Set random REP_TO in [0,REQ_TO] ¤ WHILE NOT(REP_TO expired) DO
n IF (REPLY pkt received) & flag THEN Restart REP_TO
¤ ENDWHILE ¤ Send a REPLY pkt & switch to NextRFCh
¨ IF (REPLY pkt received) & NOT flag THEN ¤ Get NextRFCh ¤ Switch to NextRFCh
AMR side SO side
19
![Page 20: a networking perspective Part 2 – option B · ATX motherboard ! 1,6 GHz Intel Pentium 4, 256 MB RAM, 160 GB HD ... Smart Object Identification ! The AMR gets a list of close by](https://reader036.vdocuments.us/reader036/viewer/2022071018/5fd1af70cd79d24a5660db8e/html5/thumbnails/20.jpg)
Extended Kalman Filter
Odometry RSSI Measures Initialization
Motes pose and robot position estimation
Localization Algorithm:
Standard SLAM 20
![Page 21: a networking perspective Part 2 – option B · ATX motherboard ! 1,6 GHz Intel Pentium 4, 256 MB RAM, 160 GB HD ... Smart Object Identification ! The AMR gets a list of close by](https://reader036.vdocuments.us/reader036/viewer/2022071018/5fd1af70cd79d24a5660db8e/html5/thumbnails/21.jpg)
Standard SLAM result
Much better than standard WSN-only localization Still large placement errors for WSNs
21
![Page 22: a networking perspective Part 2 – option B · ATX motherboard ! 1,6 GHz Intel Pentium 4, 256 MB RAM, 160 GB HD ... Smart Object Identification ! The AMR gets a list of close by](https://reader036.vdocuments.us/reader036/viewer/2022071018/5fd1af70cd79d24a5660db8e/html5/thumbnails/22.jpg)
Extended Kalman Filter
Odometry
RSSI Measures
Initialization Particle Filter
Motes pose and robot position estimation
Localization Algorithm:
SLAM with particle filter 22
![Page 23: a networking perspective Part 2 – option B · ATX motherboard ! 1,6 GHz Intel Pentium 4, 256 MB RAM, 160 GB HD ... Smart Object Identification ! The AMR gets a list of close by](https://reader036.vdocuments.us/reader036/viewer/2022071018/5fd1af70cd79d24a5660db8e/html5/thumbnails/23.jpg)
Delayed Initialization based on Particle Filter
15
23
![Page 24: a networking perspective Part 2 – option B · ATX motherboard ! 1,6 GHz Intel Pentium 4, 256 MB RAM, 160 GB HD ... Smart Object Identification ! The AMR gets a list of close by](https://reader036.vdocuments.us/reader036/viewer/2022071018/5fd1af70cd79d24a5660db8e/html5/thumbnails/24.jpg)
Step 3. Smart Objects Recognition
24
![Page 25: a networking perspective Part 2 – option B · ATX motherboard ! 1,6 GHz Intel Pentium 4, 256 MB RAM, 160 GB HD ... Smart Object Identification ! The AMR gets a list of close by](https://reader036.vdocuments.us/reader036/viewer/2022071018/5fd1af70cd79d24a5660db8e/html5/thumbnails/25.jpg)
Feature matching under motion blur (Pretto et al. ICRA 2009)
SO recognition by visual inspection
¨ The AMR should be able to recognize the objects from a description of the objects’ appearance
¨ The object appearance is stored inside the object (in the motes)
¨ The appearance is coded by scale invariant feature descriptors robust also to motion blur (MoBIF)
25
![Page 26: a networking perspective Part 2 – option B · ATX motherboard ! 1,6 GHz Intel Pentium 4, 256 MB RAM, 160 GB HD ... Smart Object Identification ! The AMR gets a list of close by](https://reader036.vdocuments.us/reader036/viewer/2022071018/5fd1af70cd79d24a5660db8e/html5/thumbnails/26.jpg)
5 meters
2 meters
1 meter
Storing the appearance
¨ Enhancing robustness to scale: ¤ the object is imaged at different distances: near, medium, far
¨ Enhancing robustness to rotation: ¤ the object is imaged every 20 deg.
¨ Limiting descriptors size ¤ MoBIF descriptors extracted at different distances and angles are merged
removing redundancy!
26
![Page 27: a networking perspective Part 2 – option B · ATX motherboard ! 1,6 GHz Intel Pentium 4, 256 MB RAM, 160 GB HD ... Smart Object Identification ! The AMR gets a list of close by](https://reader036.vdocuments.us/reader036/viewer/2022071018/5fd1af70cd79d24a5660db8e/html5/thumbnails/27.jpg)
MoBIF Matching
¨ The AMR requires the selected SO to send its MoBIF descriptors ¨ The AMR dynamically extracts MoBIF descriptor from the image taken
by onboard camera ¨ When MoBIF descriptors matching exceeds a threshold the SO is
recognized ¤ matching is robust to scale, occlusion, illumination and rotation
27
![Page 28: a networking perspective Part 2 – option B · ATX motherboard ! 1,6 GHz Intel Pentium 4, 256 MB RAM, 160 GB HD ... Smart Object Identification ! The AMR gets a list of close by](https://reader036.vdocuments.us/reader036/viewer/2022071018/5fd1af70cd79d24a5660db8e/html5/thumbnails/28.jpg)
Conclusions
¨ Synergetic combination of WSN and AMR: ¤ AMR helps WSN nodes localization ¤ WSN helps object recognition and handling by AMRs ¤ AMR & WSN together delivers new services
¨ Future work: ¤ Improve localization by using also images taken by
onboard camera into the SLAM algorithm ¤ Include smart navigation algorithms to the framework ¤ Adding object-handling algorithms to the system
28
![Page 29: a networking perspective Part 2 – option B · ATX motherboard ! 1,6 GHz Intel Pentium 4, 256 MB RAM, 160 GB HD ... Smart Object Identification ! The AMR gets a list of close by](https://reader036.vdocuments.us/reader036/viewer/2022071018/5fd1af70cd79d24a5660db8e/html5/thumbnails/29.jpg)
bibliography
¨ E. Menegatti, M. Danieletto, M. Mina, A. Pretto, A. Bardella, S. Zanconato, P. Zanuttigh, and A. Zanella, "Autonomous discovery, localization and recognition of smart objects through WSN and image features" in the proceedings of the IEEE Globecom 2010 Workshop on Towards SmArt COmmunications and Network technologies applied on Autonomous Systems, 6 december 2010, Miami, Florida, USA.
29
![Page 30: a networking perspective Part 2 – option B · ATX motherboard ! 1,6 GHz Intel Pentium 4, 256 MB RAM, 160 GB HD ... Smart Object Identification ! The AMR gets a list of close by](https://reader036.vdocuments.us/reader036/viewer/2022071018/5fd1af70cd79d24a5660db8e/html5/thumbnails/30.jpg)
Context-awareness and reactiveness
30
![Page 31: a networking perspective Part 2 – option B · ATX motherboard ! 1,6 GHz Intel Pentium 4, 256 MB RAM, 160 GB HD ... Smart Object Identification ! The AMR gets a list of close by](https://reader036.vdocuments.us/reader036/viewer/2022071018/5fd1af70cd79d24a5660db8e/html5/thumbnails/31.jpg)
Cognition-based Network
¨ Each node of the network: ¤ exploits local information to achieve its goal ¤ shares it with its neighbors
¨ Self-adaptation to the environment to achieve network wide goals
¨ Cognition applied to the entire network (not just at the PHY and MAC layers)
31
![Page 32: a networking perspective Part 2 – option B · ATX motherboard ! 1,6 GHz Intel Pentium 4, 256 MB RAM, 160 GB HD ... Smart Object Identification ! The AMR gets a list of close by](https://reader036.vdocuments.us/reader036/viewer/2022071018/5fd1af70cd79d24a5660db8e/html5/thumbnails/32.jpg)
Restricted Boltzmann Machines (RBM)
¨ Stochastic neural network ¤ input layer symmetrically connected with a hidden layer
of feature-detectors
¨ Unsupervised learning of an internal model of the data ¤ features or latent causes
32
![Page 33: a networking perspective Part 2 – option B · ATX motherboard ! 1,6 GHz Intel Pentium 4, 256 MB RAM, 160 GB HD ... Smart Object Identification ! The AMR gets a list of close by](https://reader036.vdocuments.us/reader036/viewer/2022071018/5fd1af70cd79d24a5660db8e/html5/thumbnails/33.jpg)
Hierarchical processing A key feature of cortical computation
33
![Page 34: a networking perspective Part 2 – option B · ATX motherboard ! 1,6 GHz Intel Pentium 4, 256 MB RAM, 160 GB HD ... Smart Object Identification ! The AMR gets a list of close by](https://reader036.vdocuments.us/reader036/viewer/2022071018/5fd1af70cd79d24a5660db8e/html5/thumbnails/34.jpg)
Example of learning/generation 34
![Page 35: a networking perspective Part 2 – option B · ATX motherboard ! 1,6 GHz Intel Pentium 4, 256 MB RAM, 160 GB HD ... Smart Object Identification ! The AMR gets a list of close by](https://reader036.vdocuments.us/reader036/viewer/2022071018/5fd1af70cd79d24a5660db8e/html5/thumbnails/35.jpg)
Example of applications
¨ Content-based video management ¨ Context-dependent handover in HetNets ¨ Context-aware proactive content caching ¨ …
35
![Page 36: a networking perspective Part 2 – option B · ATX motherboard ! 1,6 GHz Intel Pentium 4, 256 MB RAM, 160 GB HD ... Smart Object Identification ! The AMR gets a list of close by](https://reader036.vdocuments.us/reader036/viewer/2022071018/5fd1af70cd79d24a5660db8e/html5/thumbnails/36.jpg)
Multimedia growth
source: Cisco report (2014)
36
![Page 37: a networking perspective Part 2 – option B · ATX motherboard ! 1,6 GHz Intel Pentium 4, 256 MB RAM, 160 GB HD ... Smart Object Identification ! The AMR gets a list of close by](https://reader036.vdocuments.us/reader036/viewer/2022071018/5fd1af70cd79d24a5660db8e/html5/thumbnails/37.jpg)
Challenges for video systems
¨ Heavy data ¨ Hierarchical system ¨ Backhaul network capacity ¨ Must handle different access techniques ¨ Needs to account for video popularity,
heterogeneous user terminals ¨ Quality-of-Experience of video is hard to capture
WiFi
3G
LTE
37
![Page 38: a networking perspective Part 2 – option B · ATX motherboard ! 1,6 GHz Intel Pentium 4, 256 MB RAM, 160 GB HD ... Smart Object Identification ! The AMR gets a list of close by](https://reader036.vdocuments.us/reader036/viewer/2022071018/5fd1af70cd79d24a5660db8e/html5/thumbnails/38.jpg)
Analysis
¨ We consider a test set of 38 video clips, all encoded in an H.264-AVC format
¨ All the videos are encoded with a 16-frame structure (1 I-frame, 15 P-frames) and compressed with 18 different rates
¨ Depending on the content, the perceived quality of a compressed version changes ¤ We used the SSIM indicator to capture it
38
![Page 39: a networking perspective Part 2 – option B · ATX motherboard ! 1,6 GHz Intel Pentium 4, 256 MB RAM, 160 GB HD ... Smart Object Identification ! The AMR gets a list of close by](https://reader036.vdocuments.us/reader036/viewer/2022071018/5fd1af70cd79d24a5660db8e/html5/thumbnails/39.jpg)
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18−2.5
−2
−1.5
−1
−0.5
0
Rate versus compression
Quantization point (compression level)
ρ=lo
g(ac
tual
rate
/orig
inal
rate
)
39
![Page 40: a networking perspective Part 2 – option B · ATX motherboard ! 1,6 GHz Intel Pentium 4, 256 MB RAM, 160 GB HD ... Smart Object Identification ! The AMR gets a list of close by](https://reader036.vdocuments.us/reader036/viewer/2022071018/5fd1af70cd79d24a5660db8e/html5/thumbnails/40.jpg)
−2.5 −2 −1.5 −1 −0.5 0
0.84
0.86
0.88
0.9
0.92
0.94
0.96
0.98
1
SSIM versus rate S
elf S
imila
rity
Inde
x (S
SIM
)
ρ=log(actual rate/original rate)
40
![Page 41: a networking perspective Part 2 – option B · ATX motherboard ! 1,6 GHz Intel Pentium 4, 256 MB RAM, 160 GB HD ... Smart Object Identification ! The AMR gets a list of close by](https://reader036.vdocuments.us/reader036/viewer/2022071018/5fd1af70cd79d24a5660db8e/html5/thumbnails/41.jpg)
Remarks
¨ All the video exhibit similar trends ¤ monotonic descent ¤ a steep “fall” after a threshold
¨ However, quantitative differences mean ¤ different perceived end quality ¤ different resource requirements
¨ These characteristics are more or less consistent within the same video
41
![Page 42: a networking perspective Part 2 – option B · ATX motherboard ! 1,6 GHz Intel Pentium 4, 256 MB RAM, 160 GB HD ... Smart Object Identification ! The AMR gets a list of close by](https://reader036.vdocuments.us/reader036/viewer/2022071018/5fd1af70cd79d24a5660db8e/html5/thumbnails/42.jpg)
Requirements for video delivery
¨ QoE-based and content-aware resource allocation ¤ Rate-distortion curve depends on video content ¤ Video content affects size of the encoded video frames ¤ RBM can be used to infer rate-distortion curve of a
video by observing the size (not the content) of video frames
42
![Page 43: a networking perspective Part 2 – option B · ATX motherboard ! 1,6 GHz Intel Pentium 4, 256 MB RAM, 160 GB HD ... Smart Object Identification ! The AMR gets a list of close by](https://reader036.vdocuments.us/reader036/viewer/2022071018/5fd1af70cd79d24a5660db8e/html5/thumbnails/43.jpg)
“Our” Video Classes 43
![Page 44: a networking perspective Part 2 – option B · ATX motherboard ! 1,6 GHz Intel Pentium 4, 256 MB RAM, 160 GB HD ... Smart Object Identification ! The AMR gets a list of close by](https://reader036.vdocuments.us/reader036/viewer/2022071018/5fd1af70cd79d24a5660db8e/html5/thumbnails/44.jpg)
RBM SSIM estimate 44
0�
0,01�
0,02�
0,03�
0,04�
0,05�
1� 2� 3� 4� 5� 6� 7� 8� 9� 10�11�12�13�14�15�16�17�18�19�20�21�22�23�24�25�26�27�28�29�30�31�32�33�34�
RM
SE�
Video number�
Raw Data�
RBM�
![Page 45: a networking perspective Part 2 – option B · ATX motherboard ! 1,6 GHz Intel Pentium 4, 256 MB RAM, 160 GB HD ... Smart Object Identification ! The AMR gets a list of close by](https://reader036.vdocuments.us/reader036/viewer/2022071018/5fd1af70cd79d24a5660db8e/html5/thumbnails/45.jpg)
Cognitive video admission control
Rate-based resource allocation
Cognitive QoE-based resource allocation
Link rate over aggregate full-quality video rate
45
![Page 46: a networking perspective Part 2 – option B · ATX motherboard ! 1,6 GHz Intel Pentium 4, 256 MB RAM, 160 GB HD ... Smart Object Identification ! The AMR gets a list of close by](https://reader036.vdocuments.us/reader036/viewer/2022071018/5fd1af70cd79d24a5660db8e/html5/thumbnails/46.jpg)
Conclusions
¨ need for consolidated understanding of the fundamental limits for ¤ massive access ¤ short packets ¤ extremely variable transmission patterns
¨ Need for developing cognitive optimization modules that seamlessly interoperate at system level
46
![Page 47: a networking perspective Part 2 – option B · ATX motherboard ! 1,6 GHz Intel Pentium 4, 256 MB RAM, 160 GB HD ... Smart Object Identification ! The AMR gets a list of close by](https://reader036.vdocuments.us/reader036/viewer/2022071018/5fd1af70cd79d24a5660db8e/html5/thumbnails/47.jpg)
That’s all folks! 47