History

Next Generation Tracking

The current hard-coded filter graph implementation could be superseded by an xml-based implementation.

Enables:

  • Multi-camera support
  • Better camera calibration (OpenCV)
  • Easier configuration

Interface:

Single camera filter graph:

 1<tracker>
 2  <camera id="cam1" [...]/>
 3  <imagine inputs="cam1">
 4    <histogram id="hist" inputid="cam1" stride="3"/>
 5    <mask id="mask" inputid="camera" href="camera_mask.png"/>
 6    <distort id="distort" inputid="mask" 
 7        distortionparams="(0.69, 0)" 
 8        trapezoid="0.151537" 
 9        angle="" 
10        [...]/>
11    <crop id="crop" inputid="distort" 
12        displaydisplacement="(-174.966, -145.974)" 
13        displayscale="(2.52, 2.55)"/>
14    <historysubtract id="nohistory" inputid="crop" 
15        updateinterval="25" 
16        brighterregions="True"/>
17    <bandpass id="bandpass" inputid="nohistory" 
18        range="(1.6, 2.4)" 
19        postmult="30.0"/>
20  </imagine>
21  <blobs id="touchblobs" inputid="bandpass" 
22        threshold="148" 
23        similarity="50" 
24        areabounds="(16, 200000)" 
25        eccentricitybounds="(1, 80)"/>
26  <blobs id="trackblobs" inputid="nohistory" 
27        threshold="16" 
28        similarity="100" 
29        areabounds="(700, 200000)" 
30        eccentricitybounds="(1, 80)"/>
31</tracker>
  • Multi-camera filter graph:
    • Need to refactor pos/scale logic: Use normalized intermediate coordinates (0,0)-(1,1).
    • Graph stays mostly the same, is executed once per camera. Per-camera data:
      • camera id
      • distortion, position, crop params
  • Idea for graph:
 1<tracker>
 2  <camera id="cam1" pos="(0,0)" size="(640, 480)" 
 3      distortionparams="(0.69, 0)" 
 4      trapezoid="..." 
 5      [...]/>
 6  <camera id="cam2" [...]/>
 7  <camera id="cam3" [...]/>
 8  <imagine>
 9    [... as above, variable params are empty ...]
10  </imagine>
11  <blobs 
12   [...]
13    />
14</tracker>

Implementation:

  • Filter graph (xml repr, but no python interface)
  • Seamless cpu/gpu transition or
  • Implementation of distort, crop, historysubtract, histogram as gpu filters.
  • Single camera:
    • Call filter graph in C++
    • Adapt calibration to new filter graph
    • Refactor TrackerConfig
  • Multi-camera support:
    • Refactor pos/scale logic
    • Multi-Camera calibration
    • Custom synchronous filter graph during calibration
    • Logic for per-camera data.
    • New blob correlation step after blob finding