Advanced Tracking

Implementation

We implemented the particle tracker using the hellinger distance between color histograms and the following formula $$l(i)=e^{-\dfrac{(d_i^{HELL})^2}{2\sigma_h^2}}$$ to calculate the likelihood. The $\sigma_h$ parameter was set to $0.1$ since setting it any higher than that allowed particles far away to have a likelihood significant enough and the tracker would quickly drift if a similar object came close to the tracked one.

Based on the findings of previous assignments we used the following parameters:

• $\sigma = 1$ - Since we used the Epanechnikov kernel² we needed to determine its $\sigma$ parameter.

• $N_{bin} = 16$ - the number of bins for each color channel

We updated the template based on the parameter $\alpha$ with the formula $q_{k+1} = \alpha q_k + (1-\alpha)p_k$ where $q$ is the template histogram and $p$ is the current found object histogram.

Surprisingly a higher $\alpha=0.05$ than one used with the means-shift tracker showed better results.

Different models and parameter tuning

Lastly we used grid search to try and find the best possible set of parameters for our dataset. Because of computational reasons we used approximately five possible values for each of the parameters. Note that $\alpha = 0.05$ was fixed according to previous findings.

Additionally for RW and NCV models the $q$ was calculated based on the formula $\dfrac{min(w,h)}{6}$ where $w,h$ correspond to the patch width and height, but the value was fixed at $0.1$ for the NCA model, since any greater value allowed the particles to move everywhere across the screen. The problem with the $q$ parameter is, that if the target is moving fast a higher $q$ is needed, but a higher $q$ value also means a less stable tracker, since the particles are much more uncertain.

Note that ecause the performance of the NCA model was not good, best score of 135 failures, we only plotted the RW and NCV model performance.

We can see that a higher number of particles yields a better performance overall, but is much more computationally expensive, so we have to make a decision about the resources used and our performance.

Also, the decision of the model is very important. Since almost no entity moves with constant acceleration, the NCA model does not achieve good results. If we look at the frames, we can observe, that the particles move around radically. Looking at the basketball sequence1, when the player goes near another player the tracker is able to drift off to the other player because the particles are moving around so fiercely.

The best performance of 33 fails and 81 FPS was achieved with RW model with hundred particles. Note that the results varied, sometimes the NCV model achieved better results

Running the tracker with those parameters with the VOT evaluation toolkit³ were able to achieve an average overlap of $0.5$, $35$ failures and an average FPS of $77.03$.