Research On Small Target Detection And Tracking Method Based On Image Sequences

With the progress and development of science and technology,UAV,aerostat and other "low-small-slow(LSS)" aircrafts are becoming more and more common,which have played a huge role in promoting the development of national production industry,but also bring considerable potential threats to public security.Therefore,it is necessary to implement effective detection and stable tracking of these aircrafts.In the military field,precise monitoring is also required for non-cooperative targets with high threat,such as ballistic missiles and high-altitude unmanned reconnaissance planes.The stable detection and tracking recognition of LSS targets can be realized by using the advantages of multi-frequency optical sensors.Infrared sensors capture a target’s radiated energy to create an image.It has the advantages of long imaging distance and allweather operation,and is widely used in the early warning system and space situation awareness and other military fields.However,the existing infrared small target detection algorithm is often accompanied by a high false alarm rate when detecting targets.Visible light sensor imaging distance is close,can not be used for long-distance detection,but it has rich texture and color characteristics,has unique advantages in lowaltitude monitoring.Due to the characteristics of easy distortion and high maneuverability of LSS targets,the existing algorithms have large positioning error and low accuracy,which makes it difficult to achieve effective monitoring and stable tracking of targets.In view of the above problems,this paper proposes small target detection and tracking algorithms based on visible and infrared images.The main work is as follows:1.To solve the problem of dim and small targets detection in infrared images,an improved non-convex tensor low-rank estimation algorithm with asymmetric spatial-temporal regularization constraint is proposed.A new tensor kernel norm based on singular value reciprocal weighting is defined in the improved algorithm.Then,an adaptive weight based on structure tensor and multi-structure element top-hat is proposed to constrain the target tensor and increase the sparsity of the target tensor.Experimental results show that the improved algorithm has lower false alarm rate than the original algorithm under the same detection rate.2.A algorithm called LSS target tracking algorithm based on optical flow detection and polynomial fitting relocation an improved algorithm called visible light low-small-slow-target tracking algorithm based on improved Meanshift and were proposed for LSS targets detection and tracking in visible images.The first algorithm uses optical flow detection to obtain the suspected target area,and then uses the pre-trained SVM classifier to judge the suspected target area to eliminate the false target and determine the initial position of the target based on the continuity of target motion.In the subsequent tracking process,the relocation mechanism can be determined based on the historical movement information of the target.The relocation mechanism determines the target information by analyzing the significance of hot spots.The algorithm can effectively detect and track small targets in sky background.On the basis of the traditional Meanshift algorithm,the second algorithm uses HOG feature and gray histogram feature pair to establish a two-dimensional description template,and adds Hu invariant moment feature to establish a new convergence judgment condition.Finally,combining with the region growth method to analyze the target area change,the model update mechanism is introduced.The experiment shows that the algorithm can effectively track small aircraft targets in low altitude background.