uTrack linear motion tracker pse

作者: 凌晨2点的北京 | 来源:发表于2016-04-13 14:00 被阅读0次

uTrack.m:
- set up parameters
- trackCloseGapKalmanSparse.m:
  - Input: read the input params and stored
  - Preamble:
    - read the coordinate data
    - calculate the nearest neighbor distance for each object
    - remove the empty start and end frames
  - Link frame-to-frame:
    sequence: link forward, time reverse, link backward, time reverse, link forward
    - linkFeaturesKalmanSparse.m:
      - Initilize input
      - Reserve Memory: allocate struct memory
      - Initial the first frame info for kalman initialize: stateVec, stateCor, noiseCor
        
        first frame: kalmanInitialLinearMotion.m
        
        others: use previous kalman info
      - Link frame by frame:
        
        costMatLinearMotionLink2.m:
        
        get Input
        
        motion propagation
        
        get the transfer func, 3 kind of potential TF for linear motion
        
        For each objects in the 1st frame
        
        get previous kalman info
        
        compute new prediction
        
        store in new kalman info and will be use for next time
        
        calculate the cost matrix as the distance between prediction and actual measurement in 2rd frame
        
        then findout the minimum cost for 3 kind of TF
        
        use the user pre-defined max search radius to filter out some big distance (maxSearchRadius can also be defined by using local density)
        
        set all distance in CostMat > searchRadius to NaN and square the costMat as the upperLeft sub-matrix of the final costMat
        
        add lifetime penalty
        
        birth and death cost is set to 1.05 * CostMat.max
        
        finally construct the overall costMat as:
        
        costMat=[costMat death;birth lrblock];
        
        set all NaN in the costMat to -5
        
        use the LAP algorithm to find best solution
        
        then get the link index info:
        
        indx1c: linked objects index in frame1
        
        indx2c: linked objects index in frame2
        
        indx1u: in 1st frame but not in 2rd frame
        
        indx1c.count = indx2c.count
        
        indx1c.count + indx1u.count = total objects count in 1st frame
        
        then update the trackedFeatureInfo, near neighbor distance, prevCost matrixes
        
        all this three will be stored as [indx1c indx2c;indx1u x]
        
        update the feature life time, add 1 for miss linked objects, add 2 for linked objects
        
        calc the Kalman gain, get the best prediction and update the kalman info for next prediction to use
        
        rearrange all the tracks in the sequence of the start frame number
        
        store the tracks' each objects' [x, y, z=0, amplitude, dx, dy, dz, damp]
    - Reverse the frame sequence and repeat the process of link backward
      .
      .
      .
  - Post processing:
    - getTrackSEL: get each track's start time and end time (frame number)
    - filter out those track life time < user pre-defined
    - re-calc the nndistance

uTrack linear motion tracker pse

相关文章

网友评论

延伸阅读

深度阅读

栏目导航

热点阅读