Infrared Target Detection And Tracking Technology Based On Deep Learning Under Complex Interference Conditions

Infrared target detection and tracking technology is the core technology of infrared imaging guided missile,which has an important application prospect in the field of national defense.With the rapid development of the photoelectric countermeasure technology,the fighters of all countries are generally equipped with infrared distractors to interfere the missile from the target.Under the condition of complex interference,there will be a large number of suspected targets in the seeker’s field of view.At the same time,with different degrees of shielding of targets,it will bring severe challenges to the target detection and tracking of Infrared Air-to-Air Missile.Therefore,facing the future battlefield,the level of intelligent anti-interference needs to be improved urgently.In order to solve the problems mentioned above,focusing on the mode of “correlation filter+deep learning”,this paper studies the infrared target robustness characterization,infrared target anti-suspected disturbance,infrared target anti-occlusion and infrared target detection and tracking intelligent integration.Additionally,this paper further applies the proposed algorithm to the digital anti-interference evaluation platform,validates the effectiveness of it in engineering application and achieves the accurate detection and tracking of infrared targets.It lays a foundation for the application of intelligent algorithm in infrared imaging guided missile engineering.The main contents and innovations of this paper are as follows:1.A robust tracking method of infrared target dynamic representation is proposed,which is based on "correlation filter + deep learning" serial mode combination architecture to achieve infrared target robust tracking.This method starts from the key factors that affect the dynamic characterization of infrared target,and represents the target by improving the dynamic adaptability,discrimination ability and re-detection characterization ability.The algorithm consists of three parts: the correlation filter regression model which di stinguishes background from target,the correlation filter regression model which updates the appearance of target and the YOLOV3 deep network which is trained offline.In order to improve the recognition ability of target representation,gradient histogram feature,intensity histogram feature,brightness transform feature and color name feature are used in correlation filter regression model for fusion,and the experimental results show that the tracking performance is significantly improved.The whole tracking algorithm uses the improved average peak correlation energy to realize the adaptive updating of correlation filter and the initialization of the re-detection representation of YOLOV3 network.The results show that compared with 13 classical tracking algorithms,the tracking algorithm proposed in this chapter adopts the serial structure of “correlation filter+deep learning”,and the YOLOV3 network can re-detect and characterize the target when the tracking fails,achieving a good tracking accuracy and success rate.The experimental results show that,compared with the classical 13 tracking algorithms,the accuracy of this method is 1 and the success rate is 0.65 on the measured database of thermal imager,and the accuracy is 0.932 and the success rate is 0.528 on the public database.The optimal tracking performance is achieved on both two databases.On the simulation database under conditions of complex interference,it obtain the accuracy of 0.897 and the success rate of 0.224 respectively,which are the optimal accuracy and success rate.At the same time,the tracking speed is moderate.The causes of tracking failure are analyzed and summarized,and it is determined that suspected disturbance and various degrees of occlusion are the main causes of failure,indicating the direction of improvement.2.A stable tracking method which is characterized by fusion of infrared target depth features is proposed,and based on "correlation filter + deep learning" parallel mode combination architecture,it combines depth characteristics with previous effective features in traditional algorithms.In the case of suspected disturbance,improve the accuracy of target recognition and solve the problem of tracking point drift and jitter caused by suspected target.This method utilizes the powerful feature representation capabilities of deep convolution features.Training correlation filter regression model use traditional features and depth features,and parallel weighted fusion of response maps.According to the fusion res ponse maps,pinpoint the target location.Finally,improve the original model update mode to prevent the drift of the correlation filter regression model.When the improved average peak correlation energy is less than the predefined threshold,it will combine the initial frame model to make certain corrections.The experimental results show that,compared with 13 classical tracking algorithms,the tracking algorithm proposed in this chapter uses a parallel mode combination architecture.By introducing the depth convolution feature,when multiple infrared distractors are dropped,it obtained an accuracy of 0.948 and a success rate of 0.431 and achieved the best results.It can track infrared targets well,while the tracking speed is moderate.Compared with the algorithm in Chapter 3,the accuracy is improved by 5.1% and the success rate is improved by 20.7%.3.An anti-occlusion infrared target tracking method combining varying characteristics between frames is proposed,and based on "correlation filter+deep learning" serial and parallel mode combination architecture.It combines the inter frame variation feature with the effective feature in the traditional algorithm.In the case of occlusion,locate the target position effectively,predict the direction of target movement,and achieve anti-occlusion infrared target tracking.This method starts from the reason of tracking failure which caused by occlusion.It is determined that the spatial feature failure under the occlusion condition and center point drift problem under occlusion are the two main factors that cause the tracking failure.On the one hand,based on the LCT algorithm,obtain optical flow characteristics by using optical flow prediction network and predict the direction of target movement,The combined optical flow characteristics between frames can locate the target position effectively under occlusion conditions,According to the predicted target movement direction,it can predict the area of the target effectively in the new frame image under occlusion condition.On the other hand,when it occurs the severe deformation and occlusion,combined with feature-based matching method,using front and back frame images for sub-pixel registration to achieve tracking center point correction.The entire tracking algorithm is a joint improvement of the average peak correlation energy and the maximum response value of the judgment criteria to achieve online update of the correlation filters,and tracking center point correction.The experimental results show that,compared with 13 classical tracking algorithms,The tracking algorithm proposed in this chapter uses a combination of serial and parallel modes,by introducing optical flow network and feature point matching algorithm,it shows strong robustness when multiple infrared distractors are dropped at the same time and the target is serious or full sheltered.It obtained an accuracy of 0.966 and a success rate of0.454 respectively,while the tracking speed is moderate.Compared with the algorithm in Chapter 3,the accuracy is improved by 6.9% and the success rate is improved by 23%.4.The first intelligent integrated method of infrared target detection and tracking is proposed.According to the actual engineering application requirements of infrared imaging guided missiles under complex interference conditions,it is necessary to automatically detect aircraft target information to achieve the detection of aircraft targets of different scales,and then enter the tracking state.Intercepting and destroying incoming targets,therefore,it is necessary to intelligently integrate the detection module and the tracking module,and make full use of the advantages of each tracking module to solve the problems encountered in actual engineering.In view of the poor generalization ability of traditional detection algorithms and the inability to automatically detect aircraft targets of different scales,combined with augmented data sets and ballistic data sets for aircraft target training,based on infrared target characteristics and actual engineering needs,a suitable deep learning network is constructed.The infrared target of the initial frame is automatically detected.Aiming at the problem of target characterization and occlusion in the tracking process,combining the advantages of each chapter,the offline infrared target detection module and the online infrared target tracking module are combined.At the same time,the tracking modules are effectively complementary to achieve accurate target tracking.Finally,drawing on the experience of missile field test trial setting,combined with the digital anti-interference evaluation platform,the effectiveness of the algorithm in actual air combat confrontation is measured according to the evaluation index of the sequence image complexity proposed in Chapter 2.The experimental results show that 500 trajectories with five different sequence complexity levels are respectively selected on the constructed air combat countermeasure database.The average anti-interference probabilities of the third chapter algorithm,the fourth chapter algorithm and the fifth chapter algorithm are 81%,86.8%,and 86.2% respectively.It can be found that,compared with the algorithms in Chapter3,the algorithms in Chapters 4 and 5 play a certain role in extracting deep convolution features and optical flow features.The average anti-interference probability of the intelligent integration method in Chapter 6 is 92.2%,which is excellent anti-interference success rate.5.Aiming at the problems of insufficient high-quality and weak diversity of training samples,tracking failure in the long-term tracking process suffering complex scenarios.We propose the tracking algorithm based on adversarial fusion network.The proposed algorithm first to introduce the structure and training method of the adversarial dropout network,then proposes the adversarial spatial transformer network and the fusion method of this two networks,finally employs the adversarial fusion network to generate high-quality proposals to complete the classification of candidates and online update is achieve by the proposals.Experimental results show that the algorithm performs robust to different occlusion,deformation and fast motion.The visual tracking algorithm still has the opportunity to relocate the target when the target re-appear since the algorithm is implemented by proposals.In conclusion,this essay makes an in-depth study and analysis of infrared target detection and tracking based on deep learning under complex interference conditions and provides corresponding solutions,realized the application of the algorithm in infrared seeker,and achieved good tracking effect,which provided support for the practical engineering application of intelligent infrared imaging guidance missile.In the end of the essay,the main work is summarized,and the future research is prospected.