3. The algorithm description and goal
The algorithm`s goal is real time object tracking while
adjusting to changes in the object
The algorithm receive as input bounding box
surrounding the chosen object in the frame
The algorithm`s output is a video stream with a
bounding box around the chosen object
In addition, the algorithm supposed to be sensitive to
changes in the object`s distance and it`s angle from
the viewer
4. STRUCK algorithm flow
Receiving inputs:
𝑓𝑡, 𝑝𝑡−1, 𝑠𝑡−1
Estimating change
in object location
Calculating the
transformation on time t by
finding the most probable
transformation based on
SVM using eq.:
𝑦𝑡 = arg max 𝐹 𝑥 𝑡
𝑝 𝑡−1
, 𝑦
𝑓𝑡 = the current frame
𝑝𝑡−1= the last bbox
𝑠𝑡−1= the transformation similarity function
𝑦 = the options space for the next transformation
𝑥 𝑡 = the sampling in time t, the pattern
5. STRUCK algorithm flow
In order to find the
new bbox for time t,
the algorithm use:
𝑝𝑡 = 𝑝𝑡−1 ∘ 𝑦𝑡
Updating the
discriminant
function F with the
new position, for the
next SVM vector set
While adjusting the SVM
the algorithm uses Budget-
Maintenance function to
limit the number of vectors
for the SVM
Receiving the new
samples in time t
and checks it`s
relevance to the
object using
previous samples
Going over all the
“old” (stored)
samples and
optimizing them
Returning 𝑝𝑡, 𝑠𝑡 for
the next time frame
(for the next
calculation in time)
6. STRUCK algorithm analysis
The STRUCK approach is different from traditional
approaches by not using labels for the samples and by
predicting the next bounding box location using SVM
vector set
The algorithm`s functionality is problematic due to several
reasons:
Fast movements back and forth causes inaccuracy
The object`s learning is not exclusive to the original object
but to the bounding box
While the object exits the frame the algorithm loses it and
will recognize the first object to enter the bounding box
8. TLD algorithm analysis
The TLD algorithm has an amazing tracking and
learning ability, which characterized in the following
features:
The algorithm succeeded tracking the object even for
frequent changes in the object`s distance and visibility
While the object exits the frame the algorithm managed
to find it when it return to the framed
The algorithm keeps an extensive database for the
visibility of the object that helps it to track the object
even when it is out of the frame