An object detection strategy for uncooled infrared imagery

Abstract

An overall strategy for the very difficult problem of object detection using uncooled infrared (UCIR) sensors is discussed. The UCIR sensors are based on micro-bolometer technology and thus differ significantly from cooled infrared sensors that employ photon-counting detectors. As such, UCIR imagery tends to be very low contrast, since the sensor operates over a broad spectral band; and blurry, because of the long integration times. Ideally, the UCIR imagery would be preprocessed using an appropriate image reconstruction/restoration algorithm. If the sources of image degradation are understood and lend themselves to accurate modelling, the image reconstruction can be solved as an inverse problem. Most often this is not the case and the problem is solved using minimization approaches, such as blind deconvolution. Because image reconstruction/restoration approaches tend to be very throughput intensive, they are rarely performed in a tactical environment. More typically, a detection algorithm is applied directly to the UCIR imagery. In this paper, Local Singular Value Decomposition (LSVD) is evaluated for anomaly detection. LSVD uses local statistics to identify anomalous regions and is very good at identifying local texture differences; it appears to work quite well on UCIR imagery. Target detection results are presented for a simulated data set.