A common problem in video-based tracking of urban targets is occlusion due to buildings and vehicles. Fortunately,
when multiple video sensors are present with enough geometric diversity, track breaks due to temporary
occlusion can be substantially reduced by correlating and fusing source-level track data into system-level tracks.
Furthermore, when operating in a communication-constrained environment, it is preferable to transmit track
data rather than either raw video data or detection measurements. To avoid statistical correlation due to
common prior information, tracklets can be formed from the source tracks prior to transmission to a central
command node, which is then responsible for system track maintenance via correlation and fusion. To maximize
the operational benefit of the system-level track picture, it should be distributed in an efficient manner to all
platforms, especially the local trackers at the sensors. In this paper, we describe a centralized architecture for
multi-sensor video tracking that uses tracklet-based feedback to maintain an accurate and complete track picture
at all platforms. We will also use challenging synthetic video data to demonstrate that our architecture improves
track completeness, enhances track continuity (in the presence of occlusions), and reduces track initiation time
at the local trackers.
The central problem in multitarget, multisensor surveillance is that of determining which reports from separate sensors arise
from common objects. Due to stochastic errors in the source reports, there may be multiple data association hypotheses
with similar likelihoods. Moreover, established methods for performing data association make fundamental modeling
assumptions that hold only approximately in practice. For these reasons, it is beneficial to include some measure of
uncertainty, or ambiguity, when reporting association decisions. In this paper, we perform an analysis of the benefits
versus runtime performance of three methods of producing ambiguity estimates for data association: enumeration of the
k-best data association hypotheses, importance sampling, and Markov Chain Monte Carlo estimation. In addition, we
briefly examine the sensitivity of ambiguity estimates to violations of the stochastic model used in the data association
Fusion of data from multiple sensors can be hindered by systematic errors known as biases. Specifically, the presence of biases can lead to data misassociation and redundant tracks. Fortunately, if an estimate of the unknown biases can be obtained, the measurements and transformations for each sensor can be debiased prior to fusion. In this paper, we present an algorithm that uses targets of opportunity in the sensor field-of-view for online estimation of time-variant biases. The algorithm uses the singular value decomposition (SVD) to automatically handle the issue of parameter observability during tracking, allowing for shorter estimation windows and more accurate bias estimation. Our approach extends the novel methods proposed in the companion paper by Herman and Poore that used the SVD within a nonlinear least-squares estimator to handle the issue of parameter
observability during offine estimation of time-invariant biases using truth data.
Fusion of data from multiple sensors can be hindered by systematic errors known as biases. Specifically, the presence of biases can lead to data misassociation and redundant tracks. Fortunately, if an estimate of the unknown biases can be obtained, the measurements and transformations for each sensor can be debiased prior to fusion. In this paper, we present an algorithm that uses truth data for offline estimation of time invariant biases. Our approach is unique for two reasons. First, we explicitly avoid the use of fictitious "roll-up" biases and instead attempt to model the true sources of systematic errors. This leads to a highly nonlinear bias model that contains 18 unknown parameters. Second, we use the singular value decomposition (SVD) within our nonlinear least-squares estimator to automatically handle the issue of parameter observability. We also show how the SVD can be used to differentiate between absolute and relative bias estimates. Finally, we demonstrate that our algorithm can improve track accuracy, especially for mobile sensor platforms.
Many factors make the ground target tracking problem decidedly nonlinear and non-Gaussian. Because these factors can lead to a multimodal posterior density, a Bayesian filtering solution
is appropriate. In the last decade, the particle filter has emerged as a Bayesian inference technique that is both powerful and simple to implement. In this work, we demonstrate the necessity of using multiple-target particle filters when two or more tracks are linked through measurement contention. We also develop an efficient way to implement these filters by adaptively managing the type of particle
filters, the number of particles, and the enumeration of hypotheses during data association. Using simulated data, we compare the run-time of our adaptive particle filter algorithm to the run-times of two baseline particle filters, to demonstrate that our design mitigates the increase in computation required when performing joint
Batch maximum likelihood (ML) and maximum a posteriori (MAP) estimation with process noise is now more than thirty-five years old, and its use in multiple target tracking has long been considered to be too computationally intensive for real-time applications. While this may still be true for general usage, it is ideally suited for special needs such as bias estimation, track initiation and spawning, long-term prediction of track states, and state estimation during periods of rapidly changing target dynamics. In this paper, we examine the batch estimator formulation for several cases: nonlinear and linear models, with and without a prior state estimate (MAP vs. ML), and with and without process noise. For the nonlinear case, we show that a single pass of an extended Kalman smoother-filter over the data corresponds to a Gauss-Newton step of the corresponding nonlinear least-squares problem. Even the iterated extended Kalman filter can be viewed within this framework. For the linear case, we develop a compact least squares solution that can incorporate process noise and the prior state when available. With these new views on the batch approach, one may reconsider its usage in tracking because it provides a robust framework for the solution of the aforementioned problems. Finally, we provide some examples comparing linear batch initiation with and without process noise to show the value of the new approach.
We present a recursive Bayesian solution for the problem of joint tracking and classification of airborne targets. In our system, we allow for complications due to multiple targets, false alarms, and missed detections. More importantly, though, we utilize the full benefit of a joint approach by implementing our tracker using an aerodynamically valid flight model that requires aircraft-specific coefficients such as wing area and vehicle mass, which are provided by our classifier. A key feature that bridges the gap between tracking and classification is radar cross section (RCS). By modeling the true deterministic relationship that exists between RCS and target aspect, we are able to gain both valuable class information and an estimate of target orientation. However, the lack of a closed-form relationship between RCS and target aspect prevents us from using the Kalman filter or its variants. Instead, we rely upon a sequential Monte Carlo-based approach known as particle filtering. In addition to allowing us to include RCS as a measurement, the particle filter also simplifies the implementation of our nonlinear non-Gaussian flight model.