© saab ab 2007 multicore applications at data fusion - saab sds dr. mats ekman

15
© Saab AB 2007 Multicore Applications at Data Fusion - Saab SDS Dr. Mats Ekman

Upload: jonas-gibbs

Post on 18-Jan-2016

215 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: © Saab AB 2007 Multicore Applications at Data Fusion - Saab SDS Dr. Mats Ekman

© Saab AB 2007

Multicore Applications at Data Fusion - Saab SDS

Dr. Mats Ekman

Page 2: © Saab AB 2007 Multicore Applications at Data Fusion - Saab SDS Dr. Mats Ekman

© Saab AB 2007

Saab Data Fusion Group

• A core team of about 18 engineers, including 6 PhDs

• Active since 1984

• Air, Land, Naval, Civil domains

• Research & Development

• Marketing/Sales support

• Technical tender support

• Analysis/Design

• Implementation

• Testing, customer training Multi Sensor Tracker (MST)

Parameter tuningAlgorithm RedesignAlterations, tests

xt+1=f(xt)+wt

yt+1=h(xt)+et

Page 3: © Saab AB 2007 Multicore Applications at Data Fusion - Saab SDS Dr. Mats Ekman

03-10-06

SAAB SYSTEMS

plots

tracks

sensor

Multicore ImplementationExample 1- a success

2 step process:- get the positions- calculate scalar products and compare

with the plane

Since objects are independent parallelization of the process

TBB library (Intel Threading Building block) for C++

Page 4: © Saab AB 2007 Multicore Applications at Data Fusion - Saab SDS Dr. Mats Ekman

03-10-06

SAAB SYSTEMS

Results

Total process load

•Tested on a 4 cores local process 2.5 times faster.

•Delivered to customer - core 2.

•Drawback: need to modify the code – cannot use iterators. Some overhead using threading, cache misses?

Page 5: © Saab AB 2007 Multicore Applications at Data Fusion - Saab SDS Dr. Mats Ekman

03-10-06

SAAB SYSTEMS

Example 2 – a failure

plots

tracks

sensor

Association Process: • pre-processing – transformation to polar coordinates and clustering• Data association – work on each cluster, since cluster are independent parallelization

Technical problem:1.Static variables – several treads workingon the same variables 2. Common resources – ex. Id for tracks are obtained from a common track bank several treads in trying to access the bank lock (mute, sync)

Solution: restructure the code

Id bank

void set

Void put

Page 6: © Saab AB 2007 Multicore Applications at Data Fusion - Saab SDS Dr. Mats Ekman

03-10-06

SAAB SYSTEMS

Ongoing and Future Multicore Applicationsat Saab – CoderMP cooperation

• Particle filtering

• Anomaly detection

Page 7: © Saab AB 2007 Multicore Applications at Data Fusion - Saab SDS Dr. Mats Ekman

IDEALSIntegrated Detection and Estimation ALgorithms Solutions for data processing and fusion

Nederland

IDEALS-08:0084 1.0

7

Intro to particle filtering

A target here and now…

…expected to arrive here…

…but radar plotappeared here… …so the target is probably here

prediction – updating – prediction – updating…

Page 8: © Saab AB 2007 Multicore Applications at Data Fusion - Saab SDS Dr. Mats Ekman

IDEALSIntegrated Detection and Estimation ALgorithms Solutions for data processing and fusion

Nederland

IDEALS-08:0084 1.0

8

Probability densities

A target here and now…

…expected to arrive here…

…but radar plotappeared here…

…so the target is probably here

Page 9: © Saab AB 2007 Multicore Applications at Data Fusion - Saab SDS Dr. Mats Ekman

IDEALSIntegrated Detection and Estimation ALgorithms Solutions for data processing and fusion

Nederland

IDEALS-08:0084 1.0

9

Filtering principles

Exactly: Impractical

Ellipses/gaussian distributions: Kalman filtering

Particle filters

Page 10: © Saab AB 2007 Multicore Applications at Data Fusion - Saab SDS Dr. Mats Ekman

IDEALSIntegrated Detection and Estimation ALgorithms Solutions for data processing and fusion

Nederland

IDEALS-08:0084 1.0

10

Particle filters

Resampling

Page 11: © Saab AB 2007 Multicore Applications at Data Fusion - Saab SDS Dr. Mats Ekman

IDEALSIntegrated Detection and Estimation ALgorithms Solutions for data processing and fusion

Nederland

IDEALS-08:0084 1.0

11

Comparison (1)

Standard Kalman Constrained Kalman Particle filter

Page 12: © Saab AB 2007 Multicore Applications at Data Fusion - Saab SDS Dr. Mats Ekman

IDEALSIntegrated Detection and Estimation ALgorithms Solutions for data processing and fusion

Nederland

IDEALS-08:0084 1.0

12

Comparison (2)Particle filters - superior at severe nonlinearities

Standard Kalman Constrained Kalman Particle filter

Page 13: © Saab AB 2007 Multicore Applications at Data Fusion - Saab SDS Dr. Mats Ekman

Parallelization of PFs

Initialisation

Resampling

Normalise weights

Prediction

Update

Particle batch 1 Particle batch 2 Particle batch i Particle batch K

Initialisation

Resampling

Normalise weights

Prediction

Update

Initialisation

Resampling

Normalise weights

Prediction

Update

Initialisation

Resampling

Normalise weights

Prediction

Update

Redistribute particles between batches (i.e. communication)

Page 14: © Saab AB 2007 Multicore Applications at Data Fusion - Saab SDS Dr. Mats Ekman

Videos

Real Data from Enköping• Acoustic Sensors

• No road constraints

Simulated Data• Acoustic Sensors

• Comparison between different road constrained filters

Mix of real data from Gotland and simulated data• Radar, acoustic and seismic sensors

• Road constraints

Simulated Data• Terrain constraints

• Comparinson with only road constraints

Page 15: © Saab AB 2007 Multicore Applications at Data Fusion - Saab SDS Dr. Mats Ekman

© Saab AB 2007

Anomaly detection – complement to Rule Based Situation Assessment

Identify targets that do not behave like the majority

Here: Vessels south of Sweden.

Blue: Training data Green: Test data identified

as normal Red: Test data identified as

abnormal