|
Tasks |
|
|
|
Multiple Vector Field Model This
task aims to define the multiple vector fields (velocity and switching), as
well as the learning methods (criteria and algorithms) to estimate these
fields from video data. We will assume that we have a set of video sequences
acquired by a static surveillance camera and extract the object (pedestrians)
trajectories using a state of the art tracking algorithm. Estimation of trajectory switching probabilities The representation of trajectories as sequence of
segments of basic trajectories requires a transition model between those
basic trajectories e.g., a Markov chain, where the transition probabilities
are position dependent.. The estimation task can be formulated as a
constrained optimization problem, where probability constraints have to be
satisfied. Although general optimization techniques can be applied to deal
with this kind of problems, probability spaces have special properties that
can enrich the approach to enhance performance. Specifically, probability
spaces can be endowed with the Fisher metric. This allows the introduction of
a natural gradient with respect to this metric, and to perform optimization
using the natural gradient instead of the standard one. Model Selection Model selection questions arise in most
statistical inference/learning problems, and those to be addressed in this
project are no exception. Obviously, the main model selection problem that
will have to be faced is that of choosing the number of motion fields. The
classical overfitting/underfitting
tradeoff will arise: with an arbitrarily large number of fields we can fit
arbitrarily well the observed trajectories in the learning set, but the
generalization ability of the model will be very low, thus its performance
when used for classification or detection tasks will suffer greatly; on the
other hand, an insufficient number of fields will not be able to adequately
capture the intrinsic complexity of the scene, thus also impacting negatively
the performance of the classifiers based on the model. Human Activity Recognition – isolated pedestrians This task addresses the probelm
of estimating the motion model for two different scenarios (e.g., shopping
center and university campus) and the development of methods for activity
recognition and event detection. The detection of abnormal events can be done
using the model estimated in an unsupervised way, since we only wish to
characterize what are typical trajectories in order to detect those that are
unusual. The second goal (activity recognition) is more challenging since it
requires the specification of the activities we want to recognize (walking,
running, crossing the street, etc). Training has be
done in a supervised way, i.e., we should assign a label to each training
trajectory. Multiple People Interaction We will consider two scenarios: interaction of
pairs of pedestrians and group interaction (more than two pedestrians). We
will develop methods to determine if a pair of pedestrians know
each other or if they are strangers. We expect that people who know each
other will approach and perform some actions in a synchronized way. We will
also try to recognize joint activities such as walking together, pursuing, meeting. |