IRST System Improvements For Detection And Tracking Of DIM Moving Point Targets In Heavy Clutter

The widespread use and increasing sophistication of surveillance systems such as infrared search and track systems (IRST), both military and civilian, have generated a great deal of interest in computer algorithms capable of initiating, confirmation, developing and terminating possible targets tracks in sets of measurements. Improved infrared sensors have added infrared imagery to the type of data that can be used to detect the presence of moving targets. Unfortunately, the desire for a wide-area coverage capability in most surveillance applications often limits the amount of signal-to-noise ratio (SNR) gain most infrared systems can offer. This reduces the class of targets that can be detected and tracked to those characterized by a strong infrared signature. The desire for long-range detection, as well as the high frame rates and wide-field-of-view optics of modern staring infrared focal plane camera technology, focuses attention on algorithms to detect dim targets moving with various velocities. The detection and tracking of such infrared targets immersed within an observed scene can be difficult if the target is embedded in a dominant clutter background. It is also a major challenge for IRST systems. As one of the components has to be integrated into complete package of (Command, Control, Communication and Information) system on board, the IRST systems are wide-field-of-view surveillance systems designed for autonomous search, detection, acquisition, track and designation of potential targets. The detection capability of an IRST system is critically located in the signal processor which reduces the high-rate sensor data stream to relatively small set of potential targets per sensor frame for evaluation by a higher level track post-processor. This detection pre-processor must detect enough real targets to satisfy the system's probability of detection requirement, yet reject sufficient false targets to avoid saturating the post processor and to limit the overall system false alarm rate to an acceptable level. The feasible approaches towards addressing this issue include invoking some form of signal processing algorithms that enhances target energy while simultaneously reducing background clutter and system noise prior to thresholding and developing robust and efficient tracking algorithms. This paper focuses on developing mathematical foundations and algorithms for reliable and efficient image segmentation (clutter rejection), target detection, and target tracking based upon the assumption that IR images are composed of a mixture of component distributions, stemming from the different background and target luminance remains unchanged and targets will leave behind straight line trajectories within each processing segments of given image sequence. The principal contributions of this dissertation are the following:i. A robust hierarchical global motion estimation technique: global motion caused by camera vibration is quite common in image sequence. This nuisance factor results in translational, rotational and parallax distortions in images. So, the image registration process has become a critical and non-separable part of motion analysis and segmentation. Global motion can be modeled by a few parameters will be stated in chapter 3. These parameters estimated using regression technique which first estimate the local motion and then uses the local information to find the global motion that minimize the least square error. Then the global motion was compensated using these parameters. Refinements are carried out during the estimation inner process based on iterative elimination of singular values which introduce the bias to the estimated parameters. In order to compensate efficiently the larger image motions, the hierarchical process is introduced. The basic idea of hierarchical processing is to form a pyramid of image sequences so that on each pyramid level the motion is small enough to be estimated and the motion on the original data can be calculated as a weighted sum of results on all levels.