A number of image quality metrics (IQMs) and video quality metrics (VQMs) have been proposed in the literature for evaluating techniques and systems for mitigating degraded visual environments. Some require both pristine and corrupted imagery. Others require patterned target boards in the scene. None of these metrics relates well to the task of landing a helicopter in conditions such as a brownout dust cloud.
We have developed and used a variety of IQMs and VQMs related to the pilot’s ability to detect hazards in the scene and to maintain situational awareness. Some of these metrics can be made agnostic to sensor type. Not only are the metrics suitable for evaluating algorithm and sensor variation, they are also suitable for choosing the most cost effective solution to improve operating conditions in degraded visual environments.
Imaging sensors produce images whose primary use is to convey information to human operators. However, their proliferation has resulted in an overload of information. As a result, computational algorithms are being increasingly implemented to simplify an operator's task or to eliminate the human operator altogether. Predicting the effect of algorithms on task performance is currently cumbersome requiring estimates of the effects of an algorithm on the blurring and noise, and “shoe-horning” these effects into existing models. With the increasing use of automated algorithms with imaging sensors, a fully integrated approach is desired. While specific implementation algorithms differ, general tasks can be identified that form building blocks of a wide range of possible algorithms. Those tasks are segmentation of objects from the spatio-temporal background, object tracking over time, feature extraction, and transformation of features into human usable information. In this paper research is conducted with the purpose of developing a general performance model for segmentation algorithms based on image quality. A database of pristine imagery has been developed in which there is a wide variety of clearly defined regions with respect to shape, size, and inherent contrast. Both synthetic and “natural” images make up the database. Each image is subjected to various amounts of blur and noise. Metrics for the accuracy of segmentation have been developed and measured for each image and segmentation algorithm. Using the computed metric values and the known values of blur and noise, a model of performance for segmentation is being developed. Preliminary results are reported.
Fire fighters use relatively low cost thermal imaging cameras to locate hot spots and fire hazards in buildings. This research describes the analyses performed to study the impact of thermal image quality on fire fighter fire hazard detection task performance. Using human perception data collected by the National Institute of Standards and Technology (NIST) for fire fighters detecting hazards in a thermal image, an observer analysis was performed to quantify the sensitivity and bias of each observer. Using this analysis, the subjects were divided into three groups representing three different levels of performance. The top-performing group was used for the remainder of the modeling. Models were developed which related image quality factors such as contrast, brightness, spatial resolution, and noise to task performance probabilities. The models were fitted to the human perception data using logistic regression, as well as probit regression. Probit regression was found to yield superior fits and showed that models with not only 2nd order parameter interactions, but also 3rd order parameter interactions performed the best.
Profiling sensor systems have been shown to be effective for detecting and classifying humans against animals. A
profiling sensor with a 360 horizontal field of view was used to generate profiles of humans and animals for
classification. The sensor system contains a long wave infrared camera focused on a smooth conical mirror to
provide a 360 degree field of view. Human and animal targets were detected at 30 meters and an approximate height
to width ratio was extracted for each target. Targets were tracked for multiple frames in order to segment targets
from background. The average height to width ratio was used as a single feature for classification. The Mahalanobis
distance was calculated for each target in the single feature space to provide classification results.
Pyroelectric linear arrays can be used to generate profiles of targets. Simulations have shown that generated profiles can
be used to classify human and animal targets. A pyroelectric array system was used to collect data and classify targets as
either human or non-human in real time. The pyroelectric array system consists of a 128-element Dias 128LTI
pyroelectric linear array, an F/0.86 germanium lens, and an 18F4550 pic microcontroller for A/D conversion and
communication. The classifier used for object recognition was trained using data collected in petting zoos and tested
using data collected at the US-Mexico border in Arizona.
This paper describes the development of linear pyroelectric array systems for classification of human, animal, and
vehicle targets. The pyroelectric array is simulated to produce binary profiles of targets. The profiles are classified based
on height to width ratio using Naïve Bayesian classifiers. Profile widths of targets can vary due to the speed of the target.
Target speeds were calculated using two techniques; two array columns, and a tilted array. The profile width was
modified by the calculated speeds to show an improvement in classification results.
This paper presents initial object profile classification results using range and elevation independent features from a
simulated infrared profiling sensor. The passive infrared profiling sensor was simulated using a LWIR camera. A field
data collection effort to yield profiles of humans and animals is reported. Range and elevation independent features
based on height and width of the objects were extracted from profiles. The profile features were then used to train and
test four classification algorithms to classify objects as humans or animals. The performance of Naïve Bayesian (NB),
Naïve Bayesian with Linear Discriminant Analysis (LDA+NB), K-Nearest Neighbors (K-NN), and Support Vector
Machines (SVM) are compared based on their classification accuracy. Results indicate that for our data set SVM and
(LDA+NB) are capable of providing classification rates as high as 98.5%. For perimeter security applications where
misclassification of humans as animals (true negatives) needs to be avoided, SVM and NB provide true negative rates of
0% while maintaining overall classification rates of over 95%.