Object Tracking over Multiple Uncalibrated Cameras Using Visual, Spatial and Temporal Similarities - Advanced Concepts for Intelligent Vision Systems - page 176

Table 1 shows that the addition of location-based kernel matching improves
the overall accuracy and speed of the multi-camera tracking system, with a par-
ticularly marked increase in recall and decrease in CD. Generally, location-based
kernel matching selects a local signature better matching the foreign track sig-
nature only if the tracks are a true correspondence, i.e., only if they represent
the same real-world object. This is expected to increase true positives without
increasing false positives. Importantly, location-based kernel matching is not de-
signed to select a signature that is a worse match if the tracks are not a true
correspondence, which would result in fewer false positives. These expectations
are reflected in the relatively stronger increase in recall than precision, and no de-
crease in either. Additionally, since signature distances for true correspondences
are reduced when using location-based kernel matching, it is expected that the
condition in equation 4 will be satisfied more often and the signature weights
will be greater, thus the conditions in equation 12 will be satisfied after fewer
frames. This is reflected in the reduced CD.

Camera 1
Camera 2
Camera 3
Camera 4
Not reusing kernels
Reusing kernels

3000

2000

1000

0

0 500 1000 1500 2000 2500 3000

Frame

Fig. 3. The number of kernels stored for a test case when new kernels are created for
every observed object, and when an existing kernel is reused that is an approximation
for the size, shape and location the observed object using the method described in
section 2.1. Reusing kernels reduces memory usage and achieves better convergence.

Fig. 3 demonstrates a comparison between the number of kernels stored by
the STAC algorithm for each camera over time when kernels are created for ev-
ery observed object in each camera, and for when kernels are reused using the
method described in section 2.1. In Fig. 3, for the case of not reusing kernels, we
observe a roughly linear increase in the number of kernels stored, as expected.
In contrast, Fig. 3 shows that the number of kernels stored is bounded when
kernels are reused. These results imply a similar result for the number of his-
torical linked pairs of kernels stored. This suggests that when reusing kernels,
a guarantee can be made on the memory usage of the system. Alternatively,
a memory limit can be placed on the system without compromising tracking
accuracy. This result is particularly important for embedded applications with
restricted memory resources. Although for brevity results are shown only for one
test set, these trends are reflected in other test sets.