In this paper, we present our initial findings demonstrating a cost-effective approach to Aided Target Recognition (ATR)
employing a swarm of inexpensive Unmanned Aerial Vehicles (UAVs). We call our approach Distributed ATR (DATR).
Our paper describes the utility of DATR for autonomous UAV operations, provides an overview of our methods, and the
results of our initial simulation-based implementation and feasibility study. Our technology is aimed towards small and
micro UAVs where platform restrictions allow only a modest quality camera and limited on-board computational
capabilities. It is understood that an inexpensive sensor coupled with limited processing capability would be challenged
in deriving a high probability of detection (Pd) while maintaining a low probability of false alarms (Pfa). Our hypothesis
is that an evidential reasoning approach to fusing the observations of multiple UAVs observing approximately the same
scene can raise the Pd and lower the Pfa sufficiently in order to provide a cost-effective ATR capability. This capability
can lead to practical implementations of autonomous, coordinated, multi-UAV operations.
In our system, the live video feed from a UAV is processed by a lightweight real-time ATR algorithm. This algorithm
provides a set of possible classifications for each detected object over a possibility space defined by a set of exemplars.
The classifications for each frame within a short observation interval (a few seconds) are used to generate a belief
statement. Our system considers how many frames in the observation interval support each potential classification. A
definable function transforms the observational data into a belief value. The belief value, or opinion, represents the
UAV's belief that an object of the particular class exists in the area covered during the observation interval. The opinion
is submitted as evidence in an evidential reasoning system. Opinions from observations over the same spatial area will
have similar index values in the evidence cache. The evidential reasoning system combines observations of similar
spatial indexes, discounting older observations based upon a parameterized information aging function. We employ
Subjective Logic operations in the discounting and combination of opinions. The result is the consensus opinion from all
observations that an object of a given class exists in a given region.
This paper presents a reasoning system that pools the judgments from a set of inference agents with information
from heterogeneous sources to generate a consensus opinion that reduces uncertainty and improves knowledge
quality. The system, called Collective Agents Interpolation Integral (CAII), addresses a high level data fusion
problem by combining, in a mathematically sound manner, multi-models of inference in knowledge intensive
multi agent architecture. Two major issues are addressed in CAII. One is the ability of the inference mechanisms
to deal with hybrid data inputs from multiple information sources and map the diverse data sets to a uniform
representation in an objective space of reasoning and integration. The other is the ability of the system
architecture to allow the continuous and discrete outputs of a diverse set of inference agents to interact, cooperate,
This research is generally divided into two phases: the first phase deals with background image generation and vehicle detection, the second phase deals with vehicle tracking and video handoff.
In the first phase we view the image as a mixture of three data distributions: vehicle, background and shadow. Thus the problem is modeled as a mixture of Gaussian problem and our goal is to separate the background data from other data distributions. We proposed a median model and an improved median model to separate the background data from mixture data and to generate background reference images.
In median model we keep track of deviation between the median and its neighbors in a reordered pixel sequence. When sample size is big enough, the reordered pixel sequence is in what we called balanced-median model. This model is indicated by a very small deviation value. In this case the median of the pixel sequence falls in background set and could be used for background estimation. When sample size is not big enough, the reordered pixel sequence is in what we called shifted-median model. This model is indicated by a much bigger deviation value. In this case the median falls out of background set and are excluded for background estimation.
This median model has an impressive performance to handle slow moving or even stationary vehicles. But the time complexity is still expensive for real time image processing. The improved median model is proposed to reduce the time complexity to a reasonable level. In improved median model, we take samples in a bigger time interval to make it capable of dealing with slow moving and stationary vehicles. The sample size from experimentation is obtained as a small constant value between 5 and 20. This small sample constant size could dramatically reduce the time complexity.
As a complementary to this improved median model, a mask-classified updating method is introduced to update the background image in a short term and only classified background pixels are being used for updating.
Threshold, erosion, dilation and connected components labeling are used for noise removing and object labeling. After the first phase, the vehicle information is separated from image and input to the second phase for video hand-off and vehicle tracking. In the second phase, the weighted intensity information and shape information for each vehicle is scored and minimum-distance classification method is used for vehicle match. More than 400 vehicles are tested. An overall detection rate of 100% and tracking rate of 74% are obtained in this system.