Calibrating multiple cameras is a fundamental prerequisite for many Computer Vision applications. Typically this involves using a pair of identical synchronized industrial or high-end consumer cameras. This paper considers an application on a pair of low-cost portable cameras with different parameters that are found in smart phones. This paper addresses the issues of acquisition, detection of moving objects, dynamic camera registration and tracking of arbitrary number of targets. The acquisition of data is performed using two standard smart phone cameras and later processed using detections of moving objects in the scene. The registration of cameras onto the same world reference frame is performed using a recently developed method for camera calibration using a disparity space parameterisation and the single-cluster PHD filter.
Mahler’s Probability Hypothesis Density (PHD filter), proposed in 2000, addresses the challenges of the multipletarget
detection and tracking problem by propagating a mean density of the targets in any region of the state
space. However, when retrieving some local evidence on the target presence becomes a critical component of
a larger process - e.g. for sensor management purposes - the local target number is insufficient unless some
confidence on the estimation of the number of targets can be provided as well. In this paper, we propose a
first implementation of a PHD filter that also includes an estimation of localised variance in the target number
following each update step; we then illustrate the advantage of the PHD filter + variance on simulated data from
a multiple-target scenario.
Recent interest in multi-object filtering has focussed on the problem of discrete-time filtering, where sets of
measurements are collected at regular intervals from the sensor. Many sensors do not provide multiple measurements
at regular intervals but instead provide single-measurement reports at irregular time-steps. In this
paper we study the multi-object filtering problem for estimation from measurements where the target and clutter
processes provide measurements with Poisson arrival rates. In particular, we show that the Probability Hypothesis
Density (PHD) filter can be adapted to Poisson arrival rate measurements by modelling the probability of
detection with an exponential distribution. We demonstrate the approach in simulated scenarios.
Since the derivation of PHD filter, a number of track management schemes have been proposed to adapt the PHD filter for
determining the tracks of multiple objects. Nevertheless, the problem remains that such approaches can fail when targets
are too close or are crossing. In this paper, we propose to improve the tracking by maintaining a set of locally-based
trackers and managing the tracks with an assignment method. Furthermore, the new algorithm is based on a Gaussian
mixture implementation of the CPHD filter, by clustering neighbouring Gaussians before the update step and updating
each cluster with the CPHD filter update. In order to be computationally efficient, the algorithm includes gating techniques
for the local trackers and constructs local cardinality distributions for the targets and clutter within the gated regions. An
improvement in multi-object estimation performance has been experienced on both synthetic and real IR data scenarios.
This paper considers the effect of sensor ordering on the iterated-corrector PHD update. It is known that
changing the order of the updates results in different PHDs, however, these are usually not significantly different.
This paper considers a multisensor scenario using a single poor quality sensor in combination with good sensors,
where the bad sensor is modelled using a low probability of detection. It is shown that the quality of the updated
PHD varies significantly depending on whether the sensor is used first or last in the iterated-corrector update.
The degradation in performance of the iterated PHD filter is illustrated using a comparison of different
multisensor configurations. The OSPA error is shown to be greatest when a sensor with low probability of
detection is used in the final update of the iterated form of the PHD filter. The performance of the productmultisensor
PHD filter is also considered. The product multisensor filter is shown to perform significantly better
due to invariance to sensor ordering.
A group target is a collection of individual targets which are, for example, part of a convoy of articulated vehicles
or a crowd of football supporters and can be represented mathematically as a spatial cluster process. The process
of detecting, tracking and identifying group targets requires the estimation of the evolution of such a dynamic
spatial cluster process in time based on a sequence of partial observation sets. A suitable generalisation of the
Bayes filter for this system would provide us with an optimal (but computationally intractable) estimate of a
multi-group multi-object state based on measurements received up to the current time-step. In this paper, we
derive the first-moment approximation of the multi-group multi-target Bayes filter, inspired by the first-moment
multi-object Bayes filter derived by Mahler. Such approximations are Bayes optimal and provide estimates for
the number of clusters (groups) and their positions in the group state-space, as well as estimates for the number
of cluster components (object targets) and their positions in target state-space.
A multi-object Bayes filter analogous to the single-object Bayes filter can be derived using Finite Set Statistics for
the estimation of an unknown and randomly varying number of target states from random sets of observations.
The joint target-detection and tracking (JoTT) filter is a truncated version of the multi-object Bayes filter for
the single target detection and tracking problem. Despite the success of Finite-Set Statistics for multi-object
Bayesian filtering, the problem of multi-object smoothing with Finite Set Statistics has yet to be addressed. I
propose multi-object Bayes versions of the forward-backward and two-filter smoothers and derive optimal non-linear
forward-backward and two-filter smoothers for jointly detecting, estimating and tracking a single target
in cluttered environments. I also derive optimal Probability Hypothesis Density (PHD) smoothers, restricted to
a maximum of one target and show that these are equivalent to their Bayes filter counterparts.
The Gaussian Mixure Probability Hypothesis Density (GM-PHD) Multi-target Tracker was developed as
an extension to the GM-PHD filter to provide track continuity. The algorithm is demonstrated on forward-looking
sonar data with clutter and is compared with the results from the Particle PHD filter.