A Robust Learning-Based Detection and Tracking Algorithm - Technologies and Applications of Artificial Intelligence - page 291

Tabl e 3. Performance comparisons (pixels) of different object tracking approaches in
terms of ACE

STRUCK [1] TLD [2] Proposed Method
Crossing 51.51 16.52 6.33477

Diving 36.336 183.084 14.01

Ironman 45.3682 217.577 91.94

where O tracked
i

is overlap ratio in i th frame, R tracked
f i

is the bounding box area

from tracking algorithms and R GT
f i

is the bounding box area from ground truth.
The higher the overlap ratio, the better the performance of the video tracking
approach.
The comparison between our proposed method, STRUCK [1] and TLD [2] in
terms of overlapping ratio shown in Table 4. From the Table 4, we can conclude
that our proposed method obtain the highest average percentage for both Cross-
ing and Diving videos, while STRUCK [1] performs better in Ironman video.

Computational Time. We also compare our proposed method result with
STRUCK [1] and TLD [2] by their computational time. The comparison shown
in Table 5. From the Table 5, TLD [2] obtain the best result in all of video
sequences, while our proposed method run slower than both STRUCK [1] and
TLD [2].

Tabl e 4. Performance comparisons (percentage) of different object tracking approaches
in terms of overlap ratio

STRUCK [1] TLD [2] Proposed Method
Crossing 49.03 28.68 61.78

Diving 45.38 40.77 90.65

Ironman 20.90 8.52 8.70

Tabl e 5. Performance comparisons (seconds) different object tracking approaches in
terms of computational time

STRUCK [1] TLD [2] Proposed Method
Crossing 10.94

9.09 22.61

Diving 21.62 17.3 39.1

Ironman 21.99 16.27 39.6

3.3 Result

Our experiments aim to compare the results of the proposed approach with
existing tracking-by-detection approaches. We prove that our proposed method
can handle one of the main problems that usually occur during object tracking,
that is, occlusion problem.