Research On Infrared Small Targets Tracking Based On Improved Correlation Filter And Meta-learning

Infrared small target tracking technology has attracted attention in many fields,including early warning systems,guidance systems,and target tracking systems.Because the system needs a certain early warning capability,the system is far away from the target,which weakens the infrared radiation energy of the target.In an infrared image,the target occupies a small pixel,has no specific shape,lacks texture,is dot-like,and has only grayscale information.Therefore,how to accurately and stably track small infrared targets in complex infrared scenes with low contrast between the target and the background,insufficient appearance feature information,and susceptible to interference from background clutter,is of more research value.This paper is dedicated to improving the ability of tracking features to characterize targets from the study of infrared backgrounds.Start with model update to optimize the performance of the tracking algorithm.The main content of this paper is as follows:1)Aiming at the problems of poor anti-interference and low tracking accuracy for a single feature in complex infrared scenes such as illumination changes and background clutter.Therefore,a tracking algorithm combining multiple features is proposed.First,choose to use gray-scale features and Histogram of Oriented Gradient features for fusion.This feature can better make up for their respective disadvantages,and at the same time enrich the feature representation ability of the target,and effectively make up for the infrared dim point target tracking algorithm due to the lack of feature information.The problem of insufficient accuracy in the tracking process.In addition,according to the response value assign weights to raise the tracking accuracy of infrared small point targets.The tracking accuracy based on single feature is 89.3%,and the success rate is 54.4%.The tracking accuracy based on multi-feature fusion is92.8%,and the success rate is 67.8%.The accuracy and success rate are improved by3.5% and 13.4%,respectively.2)However,due to the method of using a fixed learning rate to update the model of each frame of image,due to the lack of an effective supervision mechanism,errors are prone to occur under the interference of external environmental factors such as move quickly and low contrast,may be used wrong samples are used to update the target model,causing tracking drift or even tracking failed.In order to improve the tracking accuracy,this paper further proposes feature fusion,through the target response value to measuring,Come to double judge the occlusion,and then control the update of the target model,so as to avoid the problem of target model pollution during occlusion;at the same time,adopt an adaptive model update method,improve the model's ability to adapt to changes in target status and complex infrared scenes.Based on the above multi-feature fusion,the tracking accuracy of the algorithm based on adaptive learning rate is 99.4%,and the success rate is 79.2%.The tracking accuracy and success rate are improved by 6.6% and 11.4%,respectively.3)In order to effectively solve the problem of insufficient training data,this paper further proposes an improved meta-learning infrared dim and small target tracking algorithm.First,the meta-learning is applied to the convolutional neural network through the pre-training tracking model,the general representation of the target are obtained through the offline training,accordingly to obtain the specific representation of the infrared dim and small target by using the initial frame target position.The target motion model is predicted by kalman filter algorithm and the optimal search area is obtained.In order to solve the problem of target loss caused by occlusion,researched the re-detection mechanism,to achieve accurate tracking of small infrared point targets.