Ask Your Question

ConDensation algo

asked 2013-09-09 04:23:09 -0500

berak gravatar image

i've been trying to come up with a 2.0 replacement for the old ConDensation algorithm that went into legacy, based on [tom sgouros patch] (

his main approach there seems to be to re-weight the random distribution in every step according to the sample-ranges, instead of keeping it fixed like in the original code.

trying this, i found, that the prediction still gets unstable, when the measurement stays still, does not change any more.

my "fix" for that was to pull the measurement into the random-distribution code, make the particles congeal, when we actually reached the target, but i have some bad feelings about it. is this the right way ? how should condension behave at all at that point ?

looking for ideas / comments on how to address the stability problem here.

if you feel like trying the original legacy code, here's a demo

the 'offending' piece of code is this (from ConDensation::updateByTime()):

if ( 0 ) // Tom's modification
    // This line may not be strictly necessary, but it prevents
    // the particles from congealing into a single particle in the
    // event of a poor choice of fitness (weighting) function.
    diff = max(diff, 0.02f * newSamples(0,d));
} else {  // my approach:
    // Rule 1 : reaching the target is the goal here, right ? 
    // * if we lost it         : swarm out  
    // * if target was reached : hog it .
    diff = min(diff, flocking * (measure(d) - newSamples(0,d)));

and here's the whole wall of code (337 lines, love it or leave it):

#include <opencv2/core/core.hpp>
#include <opencv2/core/utility.hpp>

using namespace cv;

class ConDensation 

    //! All Matrices interfaced here are expected to have dp cols, and are float.
    ConDensation( int dp, int numSamples, float flocking=0.9f );

    //! Reset. call at least once before correct()
    void initSampleSet( const Mat & lowerBound, const Mat & upperBound, const Mat & dyna=Mat() );

    //! Update the state and return prediction.
    const Mat & correct( const Mat & measurement );

    //! Access single samples(read only).
    int   sampleCount()       { return samples.rows; }
    float sample(int j,int i) { return samples(j,i); }


    int DP;                      //! Sample dimension
    int numSamples;              //! Number of the Samples                 
    float flocking;              //! flocking/congealing factor
    Mat_<float> range;           //! Scaling factor for correction, the upper bound from the orig. samples
    Mat_<float> dynamMatr;       //! Matrix of the linear Dynamics system  
    Mat_<float> samples;         //! Arr of the Sample Vectors             
    Mat_<float> newSamples;      //! Temporary array of the Sample Vectors 
    Mat_<float> confidence;      //! Confidence for each Sample            
    Mat_<float> cumulative;      //! Cumulative confidence                 
    Mat_<float> randomSample;    //! RandomVector to update sample set     
    Mat_<float> state;           //! Predicted state vector
    Mat_<float> mean;            //! Internal mean vector
    Mat_<float> measure;         //! Cached measurement vector

    struct Rand                  //! CvRandState replacement
        RNG r;
        float lo, hi;
        Rand(float l=0.0f,float h=1.0f) { set(getTickCount(),l,h); }
        void set(int64 s=0, float l=0.0f, float h=1.0f) { r.state=s; lo=l; hi=h;}
        float uni() { return r.uniform(lo,hi); }
    std::vector<Rand> rng;       //! One rng for each dimension.

    void updateByTime();

ConDensation::ConDensation( int dp, int ...
edit retag flag offensive close merge delete


Are you still maintaining this code elsewhere (github, bitbucket, etc.)? If so, can you post a link here? Has it been updated after this?

Prasanna gravatar imagePrasanna ( 2016-05-30 05:09:11 -0500 )edit

@Prasanna, code is here , if you want to play with it, but was never further updated, and personally, i've been using a KalmanFilter for any similar situation

berak gravatar imageberak ( 2016-05-30 05:20:40 -0500 )edit

I still don't get why you need to add a specific rule (swarm / hog) in the implementation of particle filter. Won't the distribution by itself take care of that? The particles might be unstable but the mean estimate of the state need not be that unstable. This blog seems to have a strategy for resampling and regularizing.

Prasanna gravatar imagePrasanna ( 2016-05-30 05:47:56 -0500 )edit

lol, i probably did not knew better, 3 years ago. ;)

(it just did not converge nicely, without)

but again, Kalman Filters ...

berak gravatar imageberak ( 2016-05-30 05:53:37 -0500 )edit

Oh. I didn't see the date :D Maybe I will try mashing it up and if something comes up I will let you know.

Prasanna gravatar imagePrasanna ( 2016-05-30 06:45:43 -0500 )edit

I found this version of Android here and works better and is more simple that the Legacy version.

I have ported to Python here

I think that must be included on the OpenCV modules again.

The link on the blog posted by Prasanna is a great reference

cesarorosco gravatar imagecesarorosco ( 2016-09-24 13:50:00 -0500 )edit

Excuse me, I have failed to implement the Particle Filter in Opencv3.3. (TrackerSamplerPF) Do you have ever successfully run the TrackerSamplerPF in Opencv3.3? Cause i can't find the official document or sample. Thanks a lot, Sincerely

Vince Li gravatar imageVince Li ( 2017-09-21 10:43:31 -0500 )edit

1 answer

Sort by ยป oldest newest most voted

answered 2016-09-24 13:51:41 -0500

edit flag offensive delete link more
Login/Signup to Answer

Question Tools


Asked: 2013-09-09 04:23:09 -0500

Seen: 1,674 times

Last updated: Sep 09 '13