Research On Detection And Tracking Algorithm Of Opto-electronical System Based On Deep Neural Network

The opto-electronical tracking system is widely used in investigation,communication,measurement,et al.Its working state includes search,capture and tracking stages,corresponding to the object detection and tracking tasks of computer vision.However,due to different application scenarios,opto-electronical tracking systems often face various problems and challenges,such as complex background interference,similar object interference,camera motion blur,poor light and strong mobility of moving objects.Therefore,to improve the overall performance of the opto-electronical tracking system,this paper focuses on the object detection and tracking tasks involved in the opto-electronical tracking system.The main contents are as follows:Firstly,the relationship between the working state of opto-electronical tracking system and the task of computer vision is studied,and the task model of computer vision is established.The search and capture state of the opto-electronical tracking system corresponds to the object detection task of computer vision,while the tracking task corresponds to the single object tracking task of computer vision.We analyze the challenges and propose a UAV detection dataset for the object detection task of opto-electronical tracking system,which lays a foundation for the follow-up work.Secondly,the diagonal network and peak response regularization technology are proposed for object detection task.Diagonal network can effectively detect UAVs,and peak response regularization method regularizes features in deep neural network.Combining plug and play with common object detection tasks,it can effectively improve the object detection performance with negligible computational cost.It also improves the performance of human pose detection and image classification tasks.Next,aiming at the challenge of target tracking task in opto-electronical tracking system,a sub peak suppression tracker is proposed.Based on the theoretical model of object tracking,two operations,sub peak pooling and boundary response truncation,are proposed to suppress the sub peak value in the tracking response map,which improve the main peak value,avoid tracking drift,and add online learning to dynamically learn the change of target.Finally,the proposed tracker achieves leading performance on six benchmarks with significant margins.Then,we analyze the online learning algorithm introduced in the object tracking task,and find that it can improve the performance but cause the speed decline,or may cause tracking drift due to learning the wrong samples.Inspired by the filtering algorithm,we model the online learning process of target tracking,and apply the filtering algorithm to process the training samples in the online learning of object tracking.Experiments are on relevant benchmarks that our method improve the performance of the tracking algorithm based on correlation filtering and deep learning.Furthermore,aiming at the efficiency of online learning in object detection task,an adaptive stochastic parallel gradient descent approach is proposed to improve the learning efficiency of online learning module in object tracking and the optimization speed of optical problems in opto-electronical tracking system,and improve the robustness of the algorithm to random disturbance parameters.Finally,the algorithm is verified in the simulation and experimental platform of optical fiber coupling task.In summary,this paper studies a series of key problems of the opto-electronical tracking system.1,For the search and capture state of the opto-electronical tracking system,diagonal network and peak response regularization module are proposed to apply to object detection task,2,UAV detection dataset is proposed.3,For the object tracking task of opto-electronical tracking system,a sub-peak suppression tracker is proposed to solve the similar target interference and background interference in the tracking process.4,For the tracking task,an online learning algorithm based on filter is proposed to solve the filtering problem of learning samples in the learning process,reduce the noise samples,and improve the tracking accuracy and speed of the tracking algorithm.5,aiming at the low efficiency of online learning in tracking task and other optical problems in opto-electronical tracking system,an adaptive stochastic parallel gradient descent approach is proposed.A large number of simulations and experiments show that the above methods can effectively improve the performance of the opto-electronical tracking system.