Multi-object Tracking Using Particle Swarm Optimization on Target Interactions - Advances in Heuristic Signal Processing and Applications - page 69

Fig. 4.4 Sub-images with object being tracked in frames #106, 107, 109, 112, 130, 140, and the
binary sub-image in frame #112

areas and height-to-width ratios in connection to location of the object at the ground
plane, is carried out to select the object candidates. For the videos that were recorded
using the calibrated cameras, we project the person locations on the ground onto 3D
world coordinates. Such 3D person's location is calculated on the basis of the center
of the bottom edge belonging to the bounding box of the blob. Then we employed
such information, together with the projected blob sizes, to enhance the delineation
of the target candidates as well as to determine the occlusions and splits of the blobs
representing the pedestrians into multiple blobs. Afterwards, the particles are ini-
tially placed in the gravity centers of the object candidates selected in such a way.
The positions of the remaining particles of the swarm are initialized on the basis of
the normal distribution which is concentrated around the state estimate at time t − 1:

x (i)
t

← N (g t − 1 ,Σ) (4.8)

where g t − 1 denotes the location of the best particle that was determined in the
previous frame at time t 1 and Σ denotes the covariance matrix of the Gaus-

sian distribution whose diagonal elements are proportional to the predicted velocity
v t =

−

g t − 2 .
In Fig. 4.4 , we can observe the behavior of the tracker with such a swarm re-
diversification. As one can notice, the tracking temporally failed in frame #109.
Thanks to placing the particles at both candidate objects (see the rightmost image on
Fig. 4.4 ), the tracker correctly recovered the identity of the person in frame #112. It
is worth noting that, owing to the object prior in the covariance matrix, the bounding
box was placed on the person undergoing tracking and not on the background areas,
see frame #109. In order to enhance the object candidate selection, we employed
also a person detector [ 11 ]. Overall, the person detector found 4550 objects in the
'S2L1_ View_1' dataset. To further enhance the re-diversification of the swarm, the
particles were initially placed on the locations determined by the person detector.

g t − 1 −

4.4 Multiple Object Tracking

The ordinary PSO is not well suited to achieve multiple object tracking. One pos-
sible approach to tackle such a problem might be to utilize a PSO that is built on
highly discriminative appearance models among different targets, for instance, like
those in [ 10 ], together with an association framework to achieve better maintaining
the identities over time. However, in practice, complex interactions between targets