Research On Infrared Small Target Detection And Tracking Algorithm Under Sky Background

Compared with the radar and laser guidance system, infrared imaging guidance technology shows the strong operational capability in the night, fog and haze weather and severe weather conditions. With the capability of high detection accuracy, strong concealment, adaptability under severe environment, infrared imaging guidance technology has been played a more and more important role in the application of military. As the key technology of infrared imaging guidance technology, infrared small target detection and tracking has been the research hotspot to the scholars all over the world.The early acquisition of the information of target is essential in the working process of infrared imaging guidance. Therefore, it is necessary to receive the radiated infrared signal in a long distance. That denotes to the long distance between target and the infrared camera. Due to the long distance, the infrared target usually manifests the characteristics of small target and exists as a spot in the image. The infrared small targets are usually blurred by the cluttered background and lack texture features, due to the radiation of the atmosphere clouds, the buildings on the ground and the noise generated by the equipment. Under the circumstance that the noise and the background with high frequency resembles the infrared small target, the detection and tracking of infrared small target is a difficult task. A lot of scholars have researched these problems and proposed some quite effective and efficient methods to detect and track infrared small target.Based on the research of predecessors, two kinds of infrared small target detection method in a single frame and detection method combined with KALMAN prediction to track infrared small target are proposed:1. According to the theory of low-rank matrix recovery, under the restrain that the residential part is sparse, the low-rank part of an image matrix can be recovered. Actually, as for the original infrared small target, the correlation degree of background pixels are strong, making sure that the background occupies the low-rank part of the image; and the correlation degree between target and noise and their surrounding pixels is weak, making them occupies the high-rank part of the image. Inspired by that, we separate the original image in to sub-blocks and arrange them in to matrixes according to the position information. Then these matrixes are decomposed separately according to ranks, which will lead to a result that the matrix background is the low-rank part and the target with some noise are the high-rank part. Then the entropy of different layers can be employed as the standard to pick out the layer with target automatically. At last the ease of noise in the target image will lead to the success of infrared small target detection.2. The amended first order directional derivatives based on facet model are employed to adjust the high-frequency part in the derivatives, and the poly-directional derivatives are acquired and fused in an effective way. Then, principle component analysis (PCA) is employed to recognize the small target, in the meantime, the two dimensional gaussian function was employed to get the training images. At last, the Euclidean distance is used to spot the attendance of the small target.3. A neighborhood blocking prediction based single frame infrared small target detection method is proposed, which is combined with KAlMAN prediction to build a method of tracking infrared small target. Where, the main frame of the single frame infrared small target detection was:the pixels in the original infrared image is replaced with the weighted neighborhood pixels. As for the background pixel, the correlation degree is strong. If they are replaced with the weighted neighborhood pixels, the difference between the replaced pixels and the original pixels are small. However, when it comes to the target pixels the correlation degree between them and their neighborhood pixels are small. If they are replaced with their weighted neighborhood pixels, the difference between the replaced pixel and the original pixels will be huge, but they will resemble the neighborhood background pixels more. Therefore, this kind of method can be employed to eliminate the target and predict the background. The important thing is to find an effective way to weigh the surrounding pixels. The weighting strategy here not only takes the correlation between the center pixel and the neighborhood pixels into account but also the correlation between the blocks with a center pixel of the predicting pixel and the ones of neighborhood pixels:the weights of the surrounding pixels are determined with the gussian-weight distance between the blocks with the center of the predicting pixel and the surrounding pixels. The smaller the distance is the bigger the weight is, which will make the details of the background well predicted. Therefore, the prominence of the small target and the acquisition of the initial position of small target in the difference image are guaranteed. For the later KALMAN prediction, the acquisition of the initial position of the target is essential. With all the data initialized, KALMAN prediction can start working. The position of the target in the next frame are predicted with KALMAN prediction, and the search area is set as a small window instead of searching in the whole frame, which makes the method easily hardware-implemented and real time-applied.