Tracking by Spatial-temporal JPDAF
In [7], a JPDAF-based approach was proposed: it formulates
the tracking problem as characterizing the position of
the moving object that maximizes appearance and motion
models. The optimal position at each time step depends
on the current appearance observations, as well as the motion
estimation obtained at the previous optimal positions.
The classical JPDAF-based tracking approach produces local
optimal solution since the decision made at time t is
based only on current measurement and previous solution
at time t − 1. If a wrong estimation of the position is selected
at time t, due to occlusions or to a bad detection, the
tracking will not be able to recover the right solution later
on.
In this section, we present a novel optimization method
for extracting the optimal path by collecting discriminating
evidences from past observations in the buffer. However, if
the detected moving regions are due to parallax, this parallax
information is propagated through tracking step, and
it interferes with the extraction of accurate trajectories of
the moving objects. In Section 4, we have proposed a re-
finement of detected moving blobs using the likelihood of a
moving pixel to belong to a moving object in the scene or
to parallax. We propose to integrate this blobs classification
into the tracking algorithm, and show that we can improve
the performance of the spatio-temporal JPDAF-based tracking
in video sequences with a strong parallax.
In the proposed JPDAF, the optimal position of the moving
object at each time step depends on several cues: the
current appearance observation, the motion estimation and
the blob’s probability to belong to a moving object or parallax.
Each cue is associated with a probability measure. We
propose to define a joint probability reflecting current and
past observations and define the appropriate data association
(i.e. tracking) by maximizing this joint probability by
collecting discriminating evidences in the sliding buffer.