Vision Tracking For Infrared Imaging Under Multi-objective Tense Scene

Target tracking is an important branch of image processing,which plays a decisive role in the construction of modern science and technology society.Tracking in infrared scenes is often used in military scenes,and infrared detection can be used to find and lock the enemy's military targets in advance.Scenario in this thesis more dense visual tracking of the infrared imaging technology,requirements for multiple targets tracking,ranging from a few to twenty,target to have the infrared small and weak state,the lack of significant color,texture and shape features,and target feature extraction is the key in the tracking,so the characteristics of weak small targets said is a big difficulty in small target tracking.On the other hand,because there are many targets to be tracked at the same time,it is necessary to ensure the stability of tracking trajectories when the target trajectories cross.The main research contents of this thesis are as follows:(1)The basic theories of common target tracking and trajectory association algorithms are studied.Firstly,the principle and characteristics of infrared imaging are described in essence.Then the correlation filtering tracking algorithm and the twin neural network algorithm are introduced.Finally,two prediction algorithms,Kalman filter and cyclic neural network,are introduced about trajectory correlation.(2)A multi-target tracking algorithm based on multi-feature correlation filtering is proposed.In order to improve the performance and effect of tracking,the correlation filter tracking framework is selected,which the feature extraction of the algorithm is improved based on.Then the Kalman filter is used to track the multi-target.On the one hand,it can prevent the target from moving too fast.On the other hand,it can effectively prevent the identity exchange in the process of target tracking combined with Hungarian matching algorithm.The final algorithm can track the target stably and prevent the exchange of target trajectories at the same time.(3)A multi-target tracking method based on siamese neural tracking network and long short-term memory network is proposed.First neural network design siamese lightweight backbone network instead of the backbone network,in addition to track prediction using long short-term memory network,through the model the modeling,the construction of the data set and model training and adjustment,compared with Kalman filter with nonlinear function,finally use Hungarian matching for same data,the correlation between the experiments prove the feasibility of the algorithm.