Experimentation
-
Changing the size of the region of interest
-
Changing the different numbers of particles (N=50, N=100, N=200, N=300)
-
Changing with RGB color histogram
-
Changing with grayscale histogram
Screenshot of the program with a small and a large window for object tracking.
Comments:
The larger the region, the longer the computation time becomes. Here, the number of errors increases.
But theoretically[2], the disadvantage, when using the color histogram, is that if a similar object is in the field of view, it risks having similar states.
In this video, the target object being a face and being the only person, there is little chance of having the same state as the target object.
Screenshot of the program with the different numbers of particles for the detection.
We can see that the more the number of particles increases, the more precise the tracking becomes.
At the beginning with 50 particles, there are some errors when the object is moving and it occludes itself. At 100 particles, the tracking becomes a little more stable. At 300, the particles are all on the skin regions.
We also notice that when it encounters large different but uniform regions, like the computer in this video, there are still some points that are not on their target with very few particles.
Screenshot of the program with a color histogram then one with grayscale.
Comment:
I can see that there is an error in my code but theoretically[1], the color histogram improves state-based tracking[1] because it gives more information unlike the gray level.
The particle filter method is a good tracking method because it is resistant to background noise and partial occultation of the tracked objects.
The color histogram allows to give more information on the object but also improves state-based tracking when used with particle filters.
Weighting allows to reduce the number of particles needed for tracking without affecting the precision, quite the contrary.
On the other hand, it is necessary to have a sufficiently large number of particles for the precision to be good, which entails a cost in terms of time calculation.
[1] ResearchGate website, article "Direction Based Modified Particle Filter for Vehicle Tracking" published in 2015
[2] ResearchGate website, article "Assessment of Vision-Based Vehicle Tracking for Traffic Monitoring Applications" , authors: Dale Joshua R. Del Carmen and Rhandley Cajote, published in 2018.