Clustering of Trajectories in Video Surveillance Using Growing Neural Gas - Foundations on Natural and Artificial Computation - page 466

Table 1. Cluster Precision for different values of maxiter and λ

maxiter

λ

3000 25000 50000 100000 300000

500 4.12 6.26 7.45 1 1

2000 NA 8.42 13.33 5.26 1

6000 NA 5.36 8.25 11.42 6.52

10000 NA 4.2 5.36 6.42 7.21

a cluster. When the number of iterations is very low and λ is also low there are
not enough nodes and their weights are not properly adapted. The extreme case
would happen when we have an only cluster. We can check that selecting these
parameters produce an adequate number of clusters. In Figure (3) it is shown
the prototype segments found with maxiter = 50000 and λ = 2000.

0

10

20

30

40

50

60

70

80

90

100

0 20 40 60 80 100

Fig. 3. Nodes found by the GNG method with maxiter = 50000 and λ = 2000

3.2 Trajectories in a Scene

In the previous section it was demonstrated how the GNG algorithm performs
appropriately for the problem of noisy segment clustering. In this section the
whole algorithm is tested within an image of the entrance lobby of the INRIA
Labs at Grenoble, filmed for the CAVIAR project with a wide angle camera. It is
necessary to say that, although the image come from real video scene application,
thetrajectorieswehaveworkedonaresynthetic trajectories. This was needed
to verify the behaviour of the method with complex trajectories and it is usually
done in most of the works for trajectory clustering. In Figure (4) some of the
synthetic trajectories are shown. It can be noticed how some of them are only
parts of completed trajectories, included to mimic the effects of occlusions and
problems with tracking algorithms.