Joint Detection, Tracking, and Classification of Multiple Targets in Clutter using the PHD Filter

Abstract

To account for joint detection, tracking, and classification (JDTC) of multiple targets from a sequence of noisy and cluttered observation sets, this paper introduces a recursive algorithm based on the probability hypothesis density (PHD) filter with the particle implementation. Assuming that each target class has a class-dependent kinematic model set, a class-matched PHD-like filter (i.e., PHD filter or its multiple-model implementation (MMPHD)) is assigned to it. In the prediction stage, the particles are propagated according to their class-dependent kinematic model set in the matched PHD-like filter. Then, the mutual information exchange between these PHD-like filters is completed by updating the particle weights in the update stage. The particles with the same class label and their corresponding weights represent the estimated class-conditioned PHD distribution. These class-conditioned PHD distributions are used to jointly estimate the number of the corresponding class targets and their states. Moreover, the algorithm incorporates the feature measurements into these PHD-like filters. The proposed multitarget JDTC algorithm has four distinctive features. First, it has a flexible modularized structure, i.e., it assigns a class-matched PHD-like filter for each target class, and facilitates the incorporation of the extra PHD-like filter for a new target class. Second, the particles can be propagated according to their exact class-dependent kinematic model set thanks to the modularized structure. Third, because of the feature measurements added and no explicit associations, it can track multiple closely spaced targets from different classes. Fourth, it avoids the possibility that the target classes with temporarily low likelihoods can end up being permanently lost. The computational burden of the proposed algorithm is linearly increased with the class number of targets. The algorithm is illustrated via a simulation example involving the tracking of two closely spaced parallel moving targets and two crossing moving targets from different classes, where targets can appear and disappear.

Development of a N-type GM-PHD filter for multiple target, multiple type visual tracking

“…Considering extensions to different target types, Yan et al [25] developed detection, tracking and classification (JDTC) of multiple targets in clutter by jointly estimating the number of targets, their kinematic states, and types of targets (classes) from a sequence of noisy and cluttered observation sets using a SMC-PHD filter. The dynamics of each target type (class) was modeled as a class-dependent model set and the signal amplitude was included in the multi-target likelihood to enhance the discrimination between targets from different classes and false alarms.…”

Multi-target tracking using a PHD-based joint tracking and classification algorithm

“…2. it can be shown that the posterior intensity can be propagated over time according to the following PHD recursion [13]. First, given the previous state ξ ∈ X , the intensities are mixed so that:…”

A Survey of Motion-Based Multitarget Tracking Methods

“…In [122], the distribution of the particles is fitted using finite mixture models (FMMs), whose parameters can be derived using a MCMC sampling scheme, then the states can be extracted according to the fitted mixture distribution. In [123], a flexible modularized structure for SMC-PHD filter is introduced, and the particles with the same class label and their corresponding weights represent the estimated class-conditioned PHD distribution. The mathematical proofs of convergence for the SMC algorithm and gives bounds for the mean-square error is presented in [124].…”