Information Technology Reference
In-Depth Information
(a)
(b)
Fig. 5.
The NIST's single camera (a) and multiple camera (b) single person tracking
MOTA evaluation medians
from the Imagery Library for Intelligent Detection Systems (i-LIDS) provided by
Home Oce Scientific Development Branch (HOSDB) in the UK [16]. We have
used the MCT video data and annotations to train and evaluate the SUNAR
performance. The data set consists of about 44 hours of video recorded by five
cameras at the London Gatwick Airport.
The task is defined as: Given 5 in situ video frames with bounding box data
specifying a person to be tracked, track the person in 5, 2 or 1 camera views by
outputting bounding boxes [20].
We have participated in the compulsory Multi-Camera Single Person Tracking
(MCSPT) and Camera Pair Single Person Tracking (CPSPT). The illustration
of the data and the area under surveillance is in figures 2, 3 and 1. For more
details see [20].
According to Johnatan Fiscus's and Martial Michels's presentation at the
2009 AVSS conference, [20] and received evaluated submissions, they used espe-
cially the Multiple Object Tracking Accuracy (MOTA, [18]) metric. The correct
detection here is when it states:
N
frames
t
=1
(
t
)
)+
c
f
(
fp
(
t
)
))
(
c
m
(
m
MOTA
=1
−
(2)
N
frames
t
=1
(
t
)
G
N
(
t
)
i
The
G
is the ground truth bounding box of an object
i
at (frame or) time
t
,
(
t
)
i
the
D
is a (SUNAR) system detection accordingly. Else the detection is false
positive
fp
)
if there is no system detection at time
t
. Then the
MOTA is defined as 2. Where
c
m
and
c
f
are weights (=1 this time) and
N
G
is
the number of ground-truth objects at time
(
t
)
, or missed
m
(
t
t
. The perfect MOTA is 1, but it
may go down to
because of false alarms [20]. The (median) MOTA results
for single camera and multiple cameras are illustrated in the figure 5. There
the camera pair run (BrnoUT 5.cpspt) was better than our multiple camera run
(BrnoUT 5.mcspt) because of the state space to be searched. Thus the single
−∞
Search WWH ::
Custom Search