The Dempster-Shafer (D-S) mass function is used in effect as a common representation of heterogeneous sensor data. In order to cast each data source in this form, first the raw data is reduced to points in a multi-dimensional feature space specific to each sensor. From there, an approach is outlined that uses a distance metric in the feature space to assign mass to each state in the class hierarchy. This hierarchy begins with the full frame of discernment which represents complete uncertainty. From there it proceeds as an n-array tree broken down into further subclasses until the finest granularity of classification for the specific sensor is reached.
For an input point to be classified, mass is assigned iteratively down the tree. In doing so, two key steps are taken. First, the uncertainty is estimated as a function of the ratio of the distance between the two closest child nodes. If the input point is deemed equidistant from the child nodes, there is a great deal of uncertainty and the mass function should reflect that. On the other hand, significant disparity indicates a much greater likelihood of one subclass. This distinction leads to the second step, where any mass not assigned to uncertainty is split between the child nodes as a function of the ratio of their distances.
The final result is a representation of the likelihood of each singleton class, as well as all unions of these classes representing uncertain states. These D-S mass functions can now be fused using Dempster's rule of combination, and classification rules can be derived to provide a more robust singular solution.
The preceding approach is derived with simulated data, and subsequently demonstrated on two sensor modalities: an ultrasonic micro-Doppler sensor and a PIR profiling sensor. The ultrasonic sensor is able to extract human motion by identifying the periodicity of a human walker's gait in the sensor field of view. The sensor can distinguish between a human, an unknown object in the scene, and background ambience. On the other hand, the profiling sensor is capable of distinguishing a horse from a human. The sensor forms a 2-D image of height versus time, and from this the orientation and eccentricity of the object are estimated and matched to known distributions of human and horse profiles. These two sensors illustrate the approach on differing hierarchies of class representations.
The Dempster-Shafer theory provides the capability of fusing orthogonal data from an ultrasonic micro-Doppler and PIR sensors. Utilizing two sets of real-world data from these sensors that were collected separately, it is possible to take a hierarchal approach to classification/discrimination through fusion of the disparate information resulting in a series of solutions with a greater confidence in comparison to a standalone sensor solution. The utilization of multiple classes afforded by the Dempster-Shafer theory increases the robustness and quality of the information from the given suite of sensors.
This work was done by Brian McGuire and Sachi Desai of the US Army RDECOM-ARDEC. ARL-0189
This Brief includes a Technical Support Package (TSP).
Dempster-Shafer Fusion for Personnel Detection
(reference ARL-0189) is currently available for download from the TSP library.
Don't have an account? Sign up here.