Research On Infrared Small And Dim Object Tracking Based On Deep Learning

Infrared small and dim target tracking is one of the core technologies of infrared search and tracking system,which is deeply concerned by scholars,and the algorithm research is very challenging.Algorithms can detect and lock moving targets in ultra-long distances and complex backgrounds.Analyzing the correlation of trajectories can significantly increase the detection rate of small and dim targets,reduce the false alarm rate,and enable the military system to make early warnings.A good infrared dim and small target detection algorithm is the cornerstone of target discovery,a robust object tracking skill is a guarantee of getting an accurate path of motion,and an accurate trajectory correlation technique is the decisive factor for judging the target.The purpose of this thesis is to do a research on the tracking and trajectory generation of infrared dim small target in cluttered background for IRST to warn early.The main works are as follow:Firstly a large amount of annotated data under the complex background of multiple scenes are collected.An infrared target dataset production process and annotation specifications are proposed.Using real infrared satellite remote sensing image as background,fitting a variety of nonlinear motion trajectories,using different kinds of real targets,man-made multi-target images are made.The production of a dataset that can be used for object detection and tracking algorithms,real and man-made data is completed,and finally an infrared dim and small target dataset containing 20,000 images and 30,000 targets is constructed.Secondly a two-scale flux density feature based on the infrared gradient vector field is proposed,which accurately distinguishes dim and small targets from complex backgrounds and noises with efficient performance and simple calculations.It is used to determine the target characteristics when trajectory correlated in the calculation of Cosine distance.Not only can ensure the efficiency of association,but also obtain a more accurate matching ability.Thirdly through research of the progress of deep learning theory in the field of object tracking in recent years,analyzing how to efficiently use deep convolutional networks to complete target tracking tasks.It is proposed to use the multi-layer feature information extracted by ResNet50 to make a longitudinal cross correlation algorithm to complete target classification and positioning task.Using the transfer learning method,an infrared dim target tracking algorithm with an accuracy rate of 0.965 and a success rate of 0.611 is trained.Finally by analyzing the data features of infrared small targets,the characteristics of tracking tasks,and the reasons that traditional tracking algorithms affect tracking performance,it is proposed to use the combination of traditional features and depth features,and use LSTM to optimize the network method by using historical frame target feature informations.Embedded the single object tracking algorithm as a prediction module to realize the trajectory correlation algorithm,and the tracking performance is improved by secondary Association,which achieves the needs of reducing false alarms,captures targets,determines targets and tracks until the targets disappear.In summary,the algorithm in this paper can effectively fulfill the requirements of early warning tasks,and the research on single object tracking algorithm and trajectory correlation algorithm achieve certain results.Under the premise of ensuring real-time performance,the single object tracking algorithm is superior to the industry tracking algorithm in the infrared small and dim target dataset and the public infrared tracking dataset.It effectively solves the problem of trajectory interruption during the target association process.Not only do this thesis ensure efficient detection rate,eliminate the false alarms of the detection algorithm,but also the number of missed objects is reduced by five times compared with the original algorithm,achieve a robust tracking effort,obtain an accuracy trajectory report.