Multi-object Tracking Using Particle Swarm Optimization on Target Interactions - Advances in Heuristic Signal Processing and Applications - page 73

is first optimized by a PSO and then by a conjugate gradient algorithm [ 26 ]. The
search area of the PSO is sufficiently large to cover promising configurations. In
the PSO, we employ 40 particles, and the maximum number of the iterations is set
to 300. The locations determined by the individual person trackers are employed
to initialize the PSO, whereas the output of the PSO is used as starting trajectory
of the conjugate gradient optimization algorithm which is responsible for the final
refinement of the trajectories. Thus, the particle swarm algorithm is utilized to seek
good local minima and the conjugate gradient is used to find the local minimum
accurately. The optimization is done using person coordinates and velocities from a
sequence of the last frames. Thus, the state vector X consists of the person locations
determined in the current frame by individual trackers and the refined locations of
all persons in a sequence of the last frames.
We achieved considerable improvement of the results by running the optimiza-
tion on only last 20 frames. For each person entering the tracking area, the opti-
mization starts in the seventh frame. In the eight frame, the optimization algorithm
runs on the current locations determined by individual trackers and the refined lo-
cations from frames #2-7, etc. Substantial improvement of the tracking accuracy
was observed in scenarios with considerable temporal occlusions. In such scenarios,
the blobs representing the pedestrians are frequently fragmented, the trackers tem-
porally loose the tracks, making uncoordinated jumps from one object to another.
Owing to the energy optimization which considers the interactions of all targets in a
sequence of the last frames, the trajectories are far smoother, and most importantly,
they pass through regions of high pedestrian likelihood.

4.5 Experiments

The algorithm was evaluated on two publicly available video sequences. The
performance of our PSO-based algorithm for multi-object tracking was com-
pared with the performance of the available PSO-based algorithm [ 30 ] for track-
ing multiple objects. In this recently proposed algorithm, species-based trackers
are employed and each tracking one object. The object interactions are mod-
eled as species competition and repulsion, whereas the occlusion is implicitly
inferred using the power of each species and the image observations. The dis-
cussed method has been evaluated on a video sequence from the PETS 2004
database which is an open database for research on visual surveillance, available
at http://homepages.inf.ed.ac.uk/rbf/CAVIAR/ . The tracking performance of our al-
gorithm was compared with the performance of the algorithm mentioned above on
an image sequence that is called 'ThreePastShop2cor', which consists of color RGB
images of size 384 288, recorded with 25 frames per second. Figure 4.5 depicts

some key frames, where three pedestrians are tracked through occlusion. All three
persons were correctly tracked in 108 frames. Thanks to patch-based representation
of the object template, the algorithm is able to select the occluding object.
The algorithm was compared with state-of-the-art algorithms for multi-object
tracking by analyses carried out both through qualitative visual evaluations as well

×