SBShankha BasuSystems research · UVA

M.S. thesis · Systems & Information Engineering

System-level modeling
of automated target
recognition.

A unified framework for characterizing recognition error, execution time, energy consumption, and hardware configuration in distributed surveillance systems.

Research overview Read the thesis

Shankha Basu
M.S., Systems & Information Engineering
University of Virginia

Thesis advisorPeter BelingCommittee chairBarry HorowitzCommittee memberMichael DeVore
αmodel
σscope
ηhardware
Performance vectorδ(α, σ, η)
accuracytimeenergy

01 A unified performance space

2003–05Core research
24,000Recognition tests
7Hardware states
121Thesis pages
01

Research question

How should an automated target-recognition system be configured when recognition accuracy, execution time, and energy consumption must be considered jointly?

The thesis addresses this question by defining independent controls for the recognition model, problem scope, and hardware platform, then mapping those controls to a common vector of system-performance measures.

Research conducted primarily 2003–2005 · Thesis submitted 2008

02

Modeling framework

A unified representation
of system performance.

Three independent control vectors map system configurations into a shared performance space.

α
Approximation control

Model

Image resolution, noise, distance and database complexity.

σ
Scope control

Recognition

Target classes, observations, assumptions and algorithm.

η
System control

Hardware

Processors, clock speed, memory, bus and bandwidth.

System performance vector

δ ( α, σ, η )
PErrRecognition error
TExecution time
EEnergy used
PAverage power
03

Experimental method

Implementation and
empirical measurement.

The experimental system combined synthetic imagery, statistical recognition, distributed software, embedded Linux, kernel-level clock control, and physical power measurement.

01

Generate

Render noisy 3D vehicle models from thousands of camera poses.

02

Model

Build statistical mean-and-mask templates at seven complexity levels.

03

Recognize

Run an integer-only likelihood classifier on embedded processors.

04

Measure

Instrument whole-node current, voltage, execution time and energy.

05

Unify

Map recognition quality and physical resource use into one performance space.

A · Recognition model

65,536 rendered views per vehicle

Camera poses were sampled across a virtual hemisphere, then aggregated into statistical templates at progressively finer levels of resolution.

B · Embedded platform

100→400MHz

Intel PXA255 nodes

Seven combinations of processor and bus frequency were exposed through a custom Linux kernel module.

C · Measurement

Whole-node power

Current was sampled at the main power supply, capturing the processor, memory, CompactFlash and wireless interface together.

Experimental scale

4vehicle classes
~4,000test images
6complexity levels
42operating points
04

Recognition results

Diminishing returns at
higher model complexity.

Increasing the number of recognition templates initially produced large accuracy gains. Near the top of the curve, computational cost kept multiplying while useful improvement nearly disappeared.

System average recognition accuracy

The curve bends sharply after k=5.

Accuracy values are averages across the four vehicle classes reported in the thesis confusion matrices.

Measured recognition accuracy

Recognition accuracy by database complexityAccuracy rises from 28.6 percent at complexity two to 99.3 percent at complexity seven, with sharply diminishing gains at the highest levels.25%50%75%100%28.6%k=252.8%k=369.1%k=492.2%k=598.7%k=699.3%k=7Database complexity
Model levelTemplates / classAccuracyExecution timeEnergy
k=51,02492.2%24.23 s67.49 J
k=64,09698.7%157.31 s417.22 J
k=716,38499.3%636.43 s1,684.68 J

Time and energy shown for system state 7, the fastest tested configuration.

Marginal tradeoff · k=6 → k=7

A marginal accuracy gain required approximately four times the measured resources.

Moving to the highest model complexity multiplied the template count by four while adding less than one percentage point of recognition accuracy.

+0.59percentage points
of accuracy
4.05×execution
time
4.04×energy
consumption
05

Energy results

Higher frequency
reduced total energy.

Lower clock speeds reduced instantaneous power but increased execution time. For this platform, the longer runtime outweighed the reduction in power draw.

State 1 · 99.5 MHz CPULower-frequency state
1,757.26 stime
2.13 Wpower
3,750.13 Jenergy
State 7 · 398.1 MHz CPUHigher-frequency state
636.43 stime
2.65 Wpower
1,684.68 Jenergy
−55%

lower total energy at the highest-frequency state, despite drawing approximately 24% more power while active.

Comparison uses the k=7 recognition workload from Tables 8.2–8.4.

06

Relation to current research

Relation to contemporary
edge-computing research.

The control-vector and performance-vector formulation concerns questions now studied in adaptive inference, hardware-aware machine learning, and edge computing. The comparison is conceptual and does not imply equivalence to contemporary neural-network systems.

01

Adaptive inference

Select just enough model complexity to satisfy the accuracy target.

02

Edge AI

Reason explicitly about energy, storage, bandwidth and latency at the device.

03

Hardware-aware ML

Evaluate an algorithm together with the machine that must execute it.

04

Runtime orchestration

Reconfigure when resources or operating requirements change.

Thesis terminology · 2008

“Choose a specific point of operation in the performance space.”

Related contemporary terminology

Adaptive inference · model routing · edge orchestration

The “then” line is a short excerpt from Chapter 1 of the thesis.

07

Scope and limitations

Contribution and
empirical boundaries.

The thesis develops a systems framework and evaluates it through an embedded implementation. The empirical study is limited to four simulated vehicle classes, one hardware family, and performance models fitted to the measured operating states.

Further validation would require real sensor data, unseen hardware configurations, systematic analysis of network scaling, and implementation of the runtime optimizer proposed in the conclusion.

Synthetic imagerySingle platform familyIn-sample model fitOptimizer as future work

Conclusion

System performance
is multidimensional.

The central contribution is a framework for evaluating recognition quality, execution time, energy consumption, and hardware controls within a common performance space.

Download the thesis Back to top ↑