Ramona Georgescu

The GM-CPHD Tracker Applied to Real and Realistic Multistatic Sonar Data Sets

The Gaussian mixture-cardinalized probability hypothesis density (GM-CPHD) tracker was applied to three real and simulated multistatic sonar data sets. The goal was to test the versatility of the tracker on data of increasing difficulty. The first two data sets presented minor challenges for the tracker and in our opinion demonstrate that it is, indeed, a tracking paradigm that is ready for “prime time.” The last data set was considerably more challenging. Without some means to use data from multiple sensors success would be in doubt. However, since a practicable multisensor form of the probability hypothesis density (PHD) filter is still unclear, predetection fusion (contact sifting) was considered a necessary first step before tracking. On all the data sets investigated, the GM-CPHD proved to be easily adaptable (e.g., to low probability of detection, large number of sensors), discovered all the targets and generated satisfactory tracks for them. Plots of the tracks obtained and the associated metrics of performance are provided.

The Multiple Model CPHD Tracker

The probability hypothesis density (PHD) is a practical approximation to the full Bayesian multi-target filter. The cardinalized PHD (CPHD) filter was proposed to deal with the “target death” problem of the PHD filter. A multiple-model PHD exists; in this work, a multiple model version of the considerably more complex CPHD filter is derived. It is implemented using Gaussian mixtures, and a track management (for display and scoring) strategy is developed.

Predetection Fusion With Doppler Measurements and Amplitude Information

In previous work, we discussed an efficient form of predetection fusion for use as a preprocessing step before tracking on data sets with large sensor networks of low-quality sensors, with particular eye to application in multistatic sonar. This 2-D version (position measurements only) was compared against an optimal (slow) technique. In this work, we present the 4-D (using position and Doppler measurements) and 5-D versions [using position, Doppler, and also aspect-dependent signal-to-noise ratio (SNR) measurements] of predetection fusion. We demonstrate that improved results, in the sense of root mean square error (RMSE) and number of declared targets-and consequently better tracking results-are possible when Doppler measurements and SNR information are incorporated into our algorithm.

Using cyclic genetic algorithms to evolve multi-loop control programs

Cyclic genetic algorithms were developed to evolve single loop control programs for robots. These programs have been used for three levels of control: individual leg movement, gait generation, and area search path finding. In all of these applications the cyclic genetic algorithm learned the cycle of actuator activations that could be continually repeated to produce the desired behavior. Although very successful for these applications, it was not applicable to control problems that required different behaviors in response to sensor inputs. Control programs for this type of behavior require multiple loops with conditional statements to regulate the branching. In this paper, we present modifications to the standard cyclic genetic algorithm that allow it to learn multi-loop control programs that can react to sensor input.

Continuous power supply for a robot colony

A continuous power supply for the robots of a colony is presented. Capacitors are used to replace the previously employed batteries, allowing more freedom of movement and providing the required continuous power. Tests were performed to determine the most effective configuration of capacitors that would allow the hexapod robot to walk for 3 minutes without recharging. Recharging of the capacitors was done through the connection of the robots metallic probes to the metal plates of a power station

Multiple model cardinalized probability hypothesis density filter

The Probability Hypothesis Density (PHD) filter propagates the first-moment approximation to the multi-target Bayesian posterior distribution while the Cardinalized PHD (CPHD) filter propagates both the posterior likelihood of (an unlabeled) target state and the posterior probability mass function of the number of targets. Extensions of the PHD filter to the multiple model (MM) framework have been published and were implemented either with a Sequential Monte Carlo or a Gaussian Mixture approach. In this work, we introduce the multiple model version of the more elaborate CPHD filter. We present the derivation of the prediction and update steps of the MMCPHD particularized for the case of two target motion models and proceed to show that in the case of a single model, the new MMCPHD equations reduce to the original CPHD equations.

Comparison of data reduction techniques based on the performance of SVM-type classifiers

In this work, we applied several techniques for data reduction to publicly available datasets with the goal of comparing how an increasing level of compression affects the performance of SVM-type classifiers. We consistently attained correct rates in the neighborhood of 90%, with the Principal Component Analysis (PCA) having a slight edge over the other data reduction methods (PLS, SRM, and OMP). One dataset proved to be hard to classify, even in the case of no dimensionality reduction. Also in this most challenging dataset, performing PCA was considered to offer some advantages over the other compression techniques. Based on our assessment, data reduction appears a useful tool that can provide a significant reduction in signal processing load with acceptable loss in performance.

Classification aided cardinalized probability hypothesis density filter

Target class measurements, if available from automatic target recognition systems, can be incorporated into multiple target tracking algorithms to improve measurement-to-track association accuracy. In this work, the performance of the classifier is modeled as a confusion matrix, whose entries are target class likelihood functions that are used to modify the update equations of the recently derived multiple models CPHD (MMCPHD) filter. The result is the new classification aided CPHD (CACPHD) filter. Simulations on multistatic sonar datasets with and without target class measurements show the advantage of including available target class information into the data association step of the CPHD filter.

The GM-CPHD applied to the corrected TNO-Blind, adjusted SEABAR07 and Metron multi-static sonar datasets

The Gaussian Mixture CardinalizedPHD (GM-CPHD) Tracker was applied to the corrected TNO-Blind dataset, the SNR adjusted datasets in SEABAR07 and to the Metron dataset generated for the MSTWG (Multistatic TrackingWorking Group). The increasing difficulty of the datasets is handled by improvements on the tracker. The tracking results (plots and metrics of performance) are included.

The CPHD and R-RANSAC trackers applied to the VIVID dataset

In this work, two multitarget trackers - the Cardinalized Probability Hypothesis Density (CPHD) filter and the Recursive Random Sample Consensus (R-RANSAC) algorithm - were applied to three scenarios of the Video Verification of IDentity (VIVID) dataset provided by DARPA. The dataset consists of real video data of multiple cars observed from an unmanned aerial vehicle (UAV) and includes challenging situations such as dense traffic and occlusions. The same detector output was given to each tracker and the same metrics of performance were computed in order to ensure fair comparison of the two tracking approaches. The results show the CPHD did better overall, which was to be expected given that it is the more mature approach.