Conclusion

This chapter explained the approach used to model fragments of behaviour. We split the original trajectory into pathlets by extracting nodes from the trajectory. These are grouped, and each group is modelled with a Gaussian spatiotemporal representation.

Several algorithms have been investigated for node selection. For the remainder of the thesis, we use algorithm A4 (page ). This algorithm, based on the mean shift algorithm, gives a better distribution of splitting nodes in the trajectory.

A spatiotemporal model of a pathlet group has been derived. This model has been extended to model the residuals as well. However, it is not clear that this extended model improves the quality of generated pathlets. In chapter 8, we assess the whole framework with and without the residual modelling.

We have developed two algorithms for grouping pathlets. We have shown visually how the greedy algorithm of section 5.4.3 outperforms the clustering of pathlets based on the normalised cut algorithm with a similarity measure based on the dynamic time warping algorithm.

The greedy algorithm is able to create pathlet groups whose models best represent the data, that is the groups they have been trained on. The approach using the normalised cuts algorithm and the dynamic time warping algorithm has more trouble creating proper pathlet models. The extracted models can generate unlikely pathlets. The main reason for that is due to the similarity measure which is not strict enough when comparing the pathlets. It can favour grouping of pathlets that only have small parts that match together.

Even if the grouping of pathlets seems to give good clusters, it does not mean that those groups are well modelled by a linear pathlet model. For grouping, the algorithm, and especially the criterion used for comparing pathlets in this algorithm, has to be adapted to the situation as we adapted the node extraction process to the model of a group.

In the remainder of the thesis, we use the greedy algorithm to extract pathlet groups from a trajectory.

In the next chapter, we describe the variable length Markov model and how it can be used to model the sequence in which each pathlet model should be visited.