Adaptive Enhancement And Motion Estimation Jointed Tracking Method For Infrared Object

With the development of photoelectric sensors,image sequences occupy a major position in the transmission and reception of information.Object tracking is helpful to dig out useful information from massive image sequences.Since infrared imaging is imaging by receiving infrared rays radiated by objects,it has the advantages of all-weather work and strong antiinterference ability,so infrared object tracking has become a hot research topic.However,complex environments,restricted imaging conditions and changing objects make infrared object tracking full of challenges.Objects in infrared image sequences often have low contrast and are not prominent enough.To solve this problem,this thesis designs a Siam RPN infrared image sequence tracking method combined with adaptive image enhancement.Furthermore,in view of the problem that the apparent modeling of the target is easily affected by factors such as occlusion and interference from similar objects,which causes the tracker to drift,this thesis is designed to combine the estimation of target motion characteristics to assist tracking.The innovative work achieved is summarized below.A Siam RPN infrared image sequences tracking method incorporating adaptive object enhancement is proposed,which is based on the Siam RPN network tracker fusion thesis to design an adaptive object enhancement module to improve the tracking accuracy.Existing image enhancement methods usually enhance the image as a whole,without considering the characteristics of the object,which often cannot meet the requirements of the tracking algorithm for the continuity of the gray distribution of the object in time and space.In this thesis,we calculate the histogram of the image grey scale and the cumulative grey scale histogram,adaptively explore the pattern of the object grey scale distribution in the image according to the scale characteristics of the object,and design a mapping function to this range to adjust the gray of the object in the image,so as to adaptively enhance the object and suppress the image background.The experimental results prove the effectiveness of the enhancement module,achieving 62.2% accuracy,35.1% robustness and 31.2% expected average overlap on the VOT-TIR2019 infrared data set.A visual object tracking method fused with Siam RPN network appearance modeling and motion characteristic estimation and correction is proposed.The method consists of four modules: apparent modeling network,motion characteristic auxiliary estimation network,model fusion and Kalman filter correction.Appearance modeling uses siamese network to extract features and uses region proposal networks to predict object positions and bounding boxes.In addition,this thesis designs a motion characteristic estimation module to improve the robustness of the occlusions,apparent changes of the object,etc.The motion estimation module uses a generative adversarial network,whose generator and discriminator are both long short term memory networks.To make the predicted trajectories smoother,this thesis also uses Kalman filtering to correct the predicted positions.The results of ablation experiments prove the effectiveness of each module.Compared with the existing infrared image sequence tracking method,the method in this thesis performs well.It has achieved63.7% accuracy,28.2% robustness and 34.5% expected average overlap on the VOTTIR2019 infrared data set.In addition,the method proposed in this thesis can be transferred to visible light tracking tasks.It has achieved 59.4% accuracy,18.7% robustness and 40.5%average expected overlap in the VOT2018 visible light data set,which demonstrates the robustness and scalability of this method.