1
Exploiting Partially Reconfigurable FPGAs for Situation-Based Reconfiguration in Wireless Sensor Networks Rafael Garcia, Dr. Ann Gordon-Ross, Dr. Alan George Target Tracking Overview & Potential of Kalman Filter Estimate state of dynamic system (moving target in this case) in a noisy environment Commonly used in target tracking, extendable for multiple targets, models, and dimensions Assume white Gaussian sensor noise Working w/ frames from USB Camera CPU-FPGA Transfer image array (1) FPGA-CPU Send X, Y, Width, Height to CPU (2) CPU Update frame Remove Backgroun d (2) Erosion 1 (1) Erosion 2 (2) Label (1) Group Propertie s (2) Sort (1) CPU Captur e frame FPGA Process image FPGA-Amenable Features Low memory/bandwidth requirements Simple filter with streaming inputs & outputs Requires only multiplication & addition, can be implemented using only logic & MACs Filter receives/produces a stream of coordinates, not images VAPRES-based processing nodes Intelligent Sensor Nodes Environment Adaptation Monitored by a process running inside the controlling agent (Microblaze) Tracking filters loaded or unloaded dynamically during run- time in response to: Number of targets inside tracking area Power available at VAPRES-based node Priority scheduling for critical targets Computation/Power Migration Hardware modules can migrate between VAPRES based processing nodes SCORES Interfac e Interfac e Interfac e Interfac e Interfac e Empty slot Empty slot Empty slot Empty slot Empty slot Switch 1 Target 1 Target 2 Target 3 Target 4 Target 5 Switch 2 Switch 3 Switch 4 Switch 5 Interfac e Interfac e Interfac e Interfac e Interfac e Sensor Interface PRR 1 Target controller Multiple targets Microblaze PRR 2 PRR 3 PRR 4 PRR 5 Display Controller Kalman filter v2 Kalman filter v1 Filtered coordinates FSL Links Network Interface Sensor Networ k Flash Controller ICAP Flash Memory (stores Kalman filter partial bitstreams) Top Level Specifications VAPRES - Virtual Architecture for Partially Reconfigurable Embedded Systems Microblaze-based reconfigurable embedded system leveraging partial reconfiguration Highly parametric architecture: e.g. number of partially reconfigurable regions (PRRs), FSL FIFOs depth, network-on-chip (SCORES) parameters Two streaming communication backbones: o Microblaze-to-coprocessors through asynchronous FSL (fast streaming links) o Intra-coprocessor communication through light-weight custom network-on-chip (SCORES) Reconfigurable module features Partial reconfigurable coprocessors can span multiple regions (PRRs) Each PRR is independently clocked o Each coprocessor can run at an independent clock frequency o Clock frequency is managed by a controlling agent (Microblaze) Reconfiguration Drivers Target behavior Target criticality Quality of sensor readings Available resources/power Situation-Based Reconfiguration Policies Characteristics Sensor type Target Type High Power High Bandwidt h Precisio n Size Kernel Type Low power Slow targets Fixed Point Small Kalman Filter Fast sampling Fast targets Fixed Point Small Kalman Filter Multi-scale Airborne targets Floating Point Large Multiscale Kalman Smoother High-Noise Filtering Noisy or generic targets Floating Point Medium Kalman Filter Kalman Filter Taxonomy Working w/ JPEG image Testbed Implementation CPU simulates the the sensor and the display

Target Tracking

  • Upload
    ronia

  • View
    37

  • Download
    2

Embed Size (px)

DESCRIPTION

Kalman filter v1. Kalman filter v2. Multiple targets. Target controller. Sensor Interface. Target 1. Target 2. Target 3. Target 4. Target 5. SCORES. Interface. Interface. Interface. Interface. Interface. Switch 5. Switch 1. Switch 2. Switch 3. Switch 4. Interface. Interface. - PowerPoint PPT Presentation

Citation preview

Page 1: Target Tracking

Exploiting Partially Reconfigurable FPGAs for Situation-Based Reconfiguration in Wireless Sensor Networks

Rafael Garcia, Dr. Ann Gordon-Ross, Dr. Alan George

Target Tracking• Overview & Potential of Kalman Filter

Estimate state of dynamic system (moving target in this case) in a noisy environment

Commonly used in target tracking, extendable for multiple targets, models, and dimensions

Assume white Gaussian sensor noise

Working w/ frames from USB Camera

CPU-FPGATransfer image

array(1)

FPGA-CPU Send X, Y, Width,

Height to CPU(2)

CPUUpdate frame

RemoveBackground

(2)

Erosion1

(1)

Erosion2

(2)

Label

(1)

GroupProperties

(2)

Sort

(1)

CPUCapture frame

FPGA Process image

• FPGA-Amenable Features Low memory/bandwidth requirements Simple filter with streaming inputs & outputs Requires only multiplication & addition, can

be implemented using only logic & MACs Filter receives/produces a stream of

coordinates, not images

VAPRES-based processing nodes

Intelligent Sensor Nodes

• Environment Adaptation Monitored by a process running inside the

controlling agent (Microblaze) Tracking filters loaded or unloaded

dynamically during run-time in response to: Number of targets inside tracking area Power available at VAPRES-based node Priority scheduling for critical targets

• Computation/Power Migration Hardware modules can migrate between

VAPRES based processing nodes Offload processing if a node runs out of

processing/power resources Migration supported by Microblaze network

interface

SCORES

Interface Interface Interface Interface Interface

Empty slot Empty slot Empty slot Empty slot Empty slot

Switch 1

Target 1 Target 2 Target 3 Target 4 Target 5

Switch 2 Switch 3 Switch 4 Switch 5

Interface Interface Interface Interface Interface

Sensor Interface

PRR

1

Target controller

Multiple targets

Microblaze

PRR

2

PRR

3

PRR

4

PRR

5

DisplayController

Kalman filter v2

Kalman filter v1

Filtered coordinatesFSL Links

NetworkInterface

SensorNetwork

FlashControllerICAP Flash Memory

(stores Kalman filter partial bitstreams)

• Top Level Specifications VAPRES - Virtual Architecture for Partially

Reconfigurable Embedded Systems Microblaze-based reconfigurable embedded

system leveraging partial reconfiguration Highly parametric architecture: e.g. number

of partially reconfigurable regions (PRRs), FSL FIFOs depth, network-on-chip (SCORES) parameters

Two streaming communication backbones:o Microblaze-to-coprocessors through asynchronous FSL

(fast streaming links)o Intra-coprocessor communication through light-weight

custom network-on-chip (SCORES)

• Reconfigurable module features Partial reconfigurable coprocessors

can span multiple regions (PRRs) Each PRR is independently clocked

o Each coprocessor can run at an independent clock frequency

o Clock frequency is managed by a controlling agent (Microblaze)

• Reconfiguration Drivers Target behavior Target criticality Quality of sensor readings Available resources/power

Situation-Based Reconfiguration Policies

CharacteristicsSensor type

Target Type High Power

High Bandwidth

Precision Size Kernel Type

Low power Slow targets Fixed Point Small Kalman FilterFast sampling Fast targets Fixed Point Small Kalman FilterMulti-scale Airborne targets Floating

PointLarge Multiscale Kalman

SmootherHigh-Noise Filtering Noisy or generic

targetsFloating

PointMedium Kalman Filter

Feature selective Multiple targets Floating Point

Large Feature Extraction/Selection

Kalman Filter Taxonomy

Working w/ JPEG image

• Testbed Implementation CPU simulates the the sensor and the display