Research On Infrared Small Target Detection And Tracking Algorithm

Infrared (IR) small target detection and tracking are key techniques in the fields of guidance, early warning and so on. In the modern wars, it is the vital factors, that detecting and tracking the enemy targets as far as possible in order to attack them at the most favorable time, which even decide the final results of the wars. The farther away from the detector, the smaller of the target size, the poorer of the image quality, and the more difficult to detect and track the targets. Therefore, the research on IR small target detection and tracking algorithms is of great significance to enhance the operating range of IR imaging system.Compared to target detection and tracking in the other area, infrared small target detection and tracking are even more difficult due to several aspects, including the low signal-to-clutter ratio, low contrast of IR image, and the small size, lack of shape and texture information of the target. Especially, the background in airborne Forward-looking infrared (FLIR) image is complex, the targets are intertwined with the clutter and there are numbers of false targets that disturb the detection and tracking. The works presented in this dissertation focus on image pre-processing, small target detection and tracking algorithm in the long-range imaging airborne FLIR image sequence under complex background.1. Research on infrared small target image pre-processing algorithmThe image pre-processing aims at suppression of the clutter in the background. The existing algorithms are mainly developed to restrain the cloud clutter and sea clutter and appear less effective in the clutter suppression under complex ground backgrounds. On the basis of analyzing the characteristics of the airborne FLIR images, A Multiple Structuring Elements Opening-by-Reconstruction Top-Hat Operator is proposed, which can effectively remove the clutter from the background. Then, according to the filtering result relying on the structuring elements, this study goes on to mark the images automatically using neural network classifier, thus avoiding the problem in the selection of structural elements, implementing the adaptive filter and further enhancing the ability to suppress the background clutter. In addition, considering the existence of certain lower contrast image after the process of filtering, a Target Energy Enhancement Algorithm Based on Wavelet Transform is given, which combines the wavelet energy image and the wavelet transform scaled image, to improve the image contrast and thus contribute to the detection of small targets.2. Research on infrared small targets detection algorithmDue to the less energy of small targets and the impact of background clutter, it is difficult to accurately detect the small targets in a single frame of FLIR images. Therefore, Multi-frame detection algorithm is typically used. However, the existing Detect-before-Track (DBT) algorithm appears less effective dealing with the images at a relatively low SCR, and the Track-before-Detect (TBD) algorithm is not suitable for real-time applications due to its large amounts of computing. According to these questions, based on image pre-processing, the kernel based Mean-Shift tracking is introduced into small target detection algorithm, a DBT algorithm is proposed: detecting the target candidates by integrating the initial multi-frame images to enhance the probability of detecting; according to the continuity of target, tracking the target candidates using Mean-Shift algorithm to remove the fake targets and reduce the false alarm; compensating the global motion of the sensor using pseudo-perspective motion model to reduce the missing of detection caused by the images shaking. Taking all these technologies above, we finally achieve the fast and accurate detection of small targets.3. Research on infrared small targets tracking algorithmIn the Model-driven Tracking algorithm, little information on the shape and texture of the small target can be used, it is difficult to set up a consummate mathematical model, and often causes the tracking failure due to delay-update or over-update of the template. In this study, a Mean-Shift Tracking algorithm based on multi-characteristics kernel density estimation is proposed. The study goes on to integrate gray-scale and the local weighted intensity entropy features to estimate the kernel density of the template and the region of target candidate, implements the tracking mission by minimizing distance between the kernel density distribution of the target candidate region and that of the template, and then automatically updates the target template according to the Bhattacharyya coefficient. This algorithm achieves the robust tracking of small targets.