Sanjay+Ranka+(PI)+ SartajSahni(CoPI) ++MarkSchmalz+(Co’PI)+€¦ · Technical+Approach+(cont’d)+ DDDAS"PIMee>ng"2"27"Jan"2016" Energy2Aware""Time"Change"Detec>on"in"SAR""2"" Ranka"(PI)"

Energy-‐Aware Time Change Detec4on Using Synthe4c Aperture Radar On High-‐Performance Heterogeneous

Architectures: A DDDAS Approach

Sanjay Ranka (PI) Sartaj Sahni (Co-‐PI) • Mark Schmalz (Co-‐PI)

University of Florida • Department of CISE

Gainesville, FL 32611-‐6120

DDDAS Program • PI Mee>ng • 27 Jan 2016

Overview of Presenta4on

1.   Research Team 2.   Technical Objec4ves 3.   Programma4c Informa4on 4.   Technical Approach & Results

•  Simula>on

•  Reconstruc>on •  Coherent Change Detec>on •  Complexity and Error Analysis

•  Energy and Power Reduc>on

5.   Ongoing and Future Work 6.   Discussion

Energy-‐Aware Time Change Detec>on in SAR -‐ Ranka (PI) DDDAS PI Mee>ng -‐ 27 Jan 2016 2

Research Team

q  Principal Inves>gator Sanjay Ranka, Ph.D. •  Research Interests: High-‐Performance Compu>ng

Energy-‐Aware Compu>ng Big Data Analy>cs

q  Co-‐PI Sartaj Sahni, Ph.D. •  Research Interests: High-‐Performance Compu>ng

Data Structures and Algorithms Signal & Image Processing

q  Co-‐PI Mark Schmalz, Ph.D, O.D. •  Research Interests: High-‐Performance Compu>ng

Signal & Image Processing Simula>on, Error Analysis

DDDAS PI Mee>ng -‐ 27 Jan 2016 3 Energy-‐Aware Time Change Detec>on in SAR -‐ Ranka (PI)

Technical Objec4ves

q  Develop Energy-‐Efficient Algorithms for Change Detec4on in Video Synthe4c Aperture Radar (SAR) Imagery q  Topics of Inves4ga4on •  Parallel Architectures (CPUs, GPUs, HMPs) •  Adap4ve Algorithms for Image Tiling •  Adap4ve Segmenta4on of SAR Pulse Dataset(s) •  Efficient SAR Image Reconstruc4on

o STEEP Constraints: Space, Time, Error, Energy Profile, and Power Consump4on •  Change Detec4on – Coherent or Incoherent?

o Effects of Noise and Clu`er o Incomplete Data o Support for Object Detec4on, Segmenta4on and Recogni4on o Complexity, Efficiency (Time, Space, and Energy or Power Consump4on)


Technical Approach

DDDAS PI Mee>ng -‐ 27 Jan 2016 5

Change detec4on accepts two input images as well as a window size, then generates a difference map.

DDDAS Approach: Output of change detector is input to algorithm that assigns spa>al resolu>on to image >les.

q  Overview of Concept

Energy-‐Aware Time Change Detec>on in SAR -‐ Ranka (PI)

Technical Approach (cont’d)


Backprojec4on is decomposed along (a) the output image dimension, where each processing device renders a >le using all the pulse data, or (b) where each processing device renders an en>re image using a subset of the pulses.

Tiling is crucial to efficient algorithm performance – the selec>on of op>mal >le size is done em-‐pirically. Pulse Data Selec4on is likewise key to re-‐construc>on algor-‐ithm efficiency. We prefetch pulses that contribute to each output (>led) pixel.




DDDAS Resolu4on and Scheduling 1.   Master decomposes the

problem into atoms (set of pulses to be rendered onto one >le of the output image).

2.   Resolu4on Controller d e t e rm i n e s s p a > a l resolu>on for render-‐ing each image >le.

3.   Master sends each atom to Mul4level Scheduler that balances load for hetero-‐geneous devices and maintains locality of access for efficiency.

q  DDDAS Approach – Mul4-‐Resolu4on



q  DDDAS Approach – Intelligent Pulse Selec4on

Four sta4c schemes for distribu4ng 100 iden4cal-‐resolu4on atoms, each with an equivalent pulse set, to 4 homogeneous processing devices.

Dynamic range dimension par44oning redistributes pulse data to minimize com-‐munica>on cost as the sensor's viewing axis moves around the scene.


1.  Simula4on •  Construct Experimental Pulse Datasets from Digital Eleva>on Maps •  Precisely Controlled Test Data •  Controllable Parameters to Facilitate Performance & Error Analysis

2.   Reconstruc4on •  Mul>resolu>on Superposi>on Algorithm Developed at UF •  Pulse Dataset Segmenta>on and Output Image Tiling for Efficiency

3.   Change Detec4on (CD) •  CD Algorithm Developed at UF Highlights Regions of Interest •  Isola>on of Regions Containing Moving Vegeta>on (high frequency ST variance) •  Construc>on of Mul>resolu>on (Pyramidal) Scene Representa>on •  Iden>fica>on of Future Target Movement Regions (by variance & context) •  Sta>s>cal ID and Tracking of Targets by Spa>otemporal (ST) Variance



4.   Complexity and Error Analysis (sta%c : compile-‐>me and dynamic : run-‐>me)


HIDEF Processing Algorithm HIDEF Algorithm Data Structures Functional Analysis Data Structure Layout Computational Analysis Memory Analysis Error Propagation Analysis I/O Channel Analysis

Architectural Constraints

Time-Space-Error Cost Analysis

Array indexing - row-major - column-major - custom scan

Partitioning/Connectivity Interleaving/Caching Bandwidth/Error rate

Packet size & overhead Bandwidth / Connectivity Error characteristics

Algorithm structure Procedure calls Operation mix

Execution trace Complexity Work estimate

Forward analysis Backward analysis Uncertainty quantifi-cation

VSAR VSAR



5.   Energy and Power Consump4on Reduc4on •  Basis: (1) Algorithmic Model + (2) Energy/Power Measurements

• Mul4resolu4on Structure Facilitates Region Segmenta4on

! Non-‐Target Regions Reconstructed at Low Resolu4on o Example: Foliage and cover regions o  Increased computa>onal efficiency o Decreased power consump>on

! Probable Target Regions Reconstructed at Higher Resolu4on o Detail preserved to facilitate target recogni>on and tracking o  Increases success of change detec>on algorithm applica>on o Target predic>on model (speed, direc>on) predicts range of posi>ons in next frame o Energy and power consump>on models constrain processing strategy for next frame o Result: Data-‐ and processing-‐directed reduc>on of power consump>on



q  Results: Simula4on •  Construct Experimental Pulse Datasets from Digital Eleva4on Maps


x

y

z Pulse Emi`er Receiver

ζsr S

T

Given: Bidirec>onal Reflec>vity Distribu>on Func>on (BRDF) Emiher Intensity I Received Intensity Ir = I • BRDF(θi) Emiher Track S = (s1, s2, …, sP) S = T ⇒ monostatic SAR Receiver Track T = (t1, t2, …, tP) S ≠ T ⇒ bistatic SAR Number of Pulses P Pulse Data Resolu>on (per pulse) NB bins Resolu>on NxN pixels of Reconstructed Image defined on X Frame Rate F

Objec@ve: Construct pulse dataset D having NP pulses of NB bins each → F⋅ P ⋅ NB elements

Simulated Ground Plane X



q  Results: Simula4on & Reconstruc4on Example 1.  Construct Digital Eleva>on Map 2.  Construct Experimental Pulse Datasets from Digital Eleva>on Maps 3.  Compute DEM from Pulse Dataset using SAR Reconstruc>on Algorithm


Image of DEM Monosta4c SAR Reconstruc4on with Flare Ar4facts

Early Simula4on Results (Reconstruc4on) ► ζsr = 45 degrees ► Emiher alt = 7,071m ► Receiver alt = 106m ► Lateral separa>on of central receiver from transmiher = 9,850m ► 600x600 pixel image



q  Results: Preliminary Reconstruc4on Timing



q  Results: Preliminary Reconstruc4on Timing (May 2015, cont’d)


28.4 sec / frame (40962 pixels) GPU exe 4me


q  Results: Current Algorithm Testbed


Master Processor: •  CPU: Intel Xeon clocked at 2.0 GHz. •  Opera>ng system : Ubuntu 12.04.4 LTS SMP

Co-‐Processor: •  GPU: Nvidia Tesla C2050, 1 GPU processor with 448 CUDA cores,

clocked at 1.147 GHz •  GPU Memory 6GB RAM clocked at 1.5 GHz

Current Reconstruc4on Timing Results (Jan 2016): •  Single GPU latency of 2.72 sec / frame [40962 pixels] •  Speedup (versus May 2015) = 28.4 sec / 2.72 sec = 10.4X


q  Results: Reconstruc4on Timing, One Nvidia FermiTM GPU •  Measure Algorithm Performance with & without External Overhead


TileSz = 4 TileSz = 8 TileSz = 16 TileSz = 32

Execu4

on + I/O Tim

e, se

c

512  1024 2048 4096 8192 Image Dimension, pixels

512  1024 2048 4096 8192 Image Dimension, pixels

Execu4

on + I/O Tim

e, se

c

TileSz = 4 TileSz = 8 TileSz = 16 TileSz = 32

Result: Op4mal Tile Size = 16 pixels


2.72 sec / frm (40962 pixels) GPU exe 4me


q  Results: Reconstruc4on Timing, Four Nvidia FermiTM GPUs •  Measure Algorithm Performance with & without Charm++


Charm++ Slower: Why? -‐-‐ 100ms setup >me -‐-‐ Lacks automa>c data par>>oning that we have -‐-‐ Our approach op>m-‐ ized for SAR -‐-‐ Charm performance improves with data size


q  Results: Preliminary Change Detec4on Algorithm


1: 250m x 170m road scene (SAR) 2: Car traverses road from right to lep

3: Mean of 25 previous frames 4: Standard devia>on of 25 previous frames


q  Results: Preliminary Change Detec4on Algorithm (cont’d)


5: Compute z-‐score per pixel 6: Apply 15x15-‐pixel window filter

7: Threshold by mean and st-‐dev 8: Apply algorithm framewise, per >le

False Posi4ve due to tree blowing in the wind.

Increase Efficiency by focusing on hi-‐resolu>on areas (probable target) and contextual cues (e.g., car on road)

Preliminary Conclusions & Future Work

q  Accomplishments ü  Improved Reconstruc4on Algorithm > 10X Faster than Previous ü  Preliminary Change Detec4on Algorithm Developed & Tested ü  SAR Simulator Developed for Precise Control in Test & Analysis ü  Publica4ons (prepared in conjunc>on with this project):


John Wiley & Sons, Inc.

" Monograph: preliminary version sub-‐ mihed to John Wiley & Sons

Journal Paper ! Mul4sta4c SAR submihed to IEEE Tr. AES (in revision)

Preliminary Conclusions & Future Work

q  Future Work (next 12 months)

Ø  Performance Op4miza4on: Reconstruc>on & Change Detec>on (RCD) Algorithms

Ø  Improve CD Algorithm Performance: á Detec>on Probability, â False Alarms

Ø  Enhancement of Mul4resolu4on Structures to Focus Processing for Efficiency o  More Accurately Determine Probable Target Regions (High Resolu>on) o  Increase False Alarm Rejec>on via Temporal Spectral Filtering (Low Resolu>on)

Ø  Measure & Improve: RCD Algorithms’ Energy Profiles and Power Consump>on

Ø  Enhance SAR/VSAR Simulator with Flare, Interference, & Diffrac>on Effects

Ø  Simula4on/Test/Analysis of RCD Algorithms for Distributed Networked Embedded Apps


Discussion


Documents

Sanjay+Ranka+(PI)+ SartajSahni(CoPI) ++MarkSchmalz+(Co’PI)+€¦ · Technical+Approach+(cont’d)+ DDDAS"PIMee>ng"2"27"Jan"2016" Energy2Aware""Time"Change"Detec>on"in"SAR""2"" Ranka"(PI)"