Research On High Precision Feature Positioning And Tracking Technology Of Dynamic Extended Object

In the electro-optical tracking system based on extended target imaging,the localization of the target image will have a great impact on the subsequent tracking control performance.How to achieve high-precision feature localization from a series of target sequence images with complex background and large extension has been a very active research direction.The target image studied in this paper has a certain degree of expansion,texture and shape information,but the moving process of the target is more complex,and the scale change,rotation,occlusion and other situations may occur,which will lead to the decline of tracking accuracy.Therefore,it is necessary to study a high robustness feature localization algorithm for extended target images.Feature expression is a key part of the target tracking process.The artificial features are relatively simple and real-time,but there is a problem of insufficient representation ability.It is easy to drift when dealing with problems such as rapid change and target occlusion.With the strong feature expression ability of deep neural networks in target detection and recognition tasks,deep neural network features are gradually used as feature extraction tools,but how to use and integrate these features is still worth studying.Based on the consideration of real-time target tracking,this paper takes correlation filtering as the basic framework to study the algorithm from the feature improvement route.In the field of object detection and recognition,the increase of the number of feature layers is positive gain for the algorithm effect.This paper takes the residual neural network(Resnet-50)with deeper number of feature layers than the VGG-19 commonly used at present as the main research object,and analyzes in detail the influence of each feature layer on the target tracking performance.Innovatively,feature fusion of ResNet-50 special structures-additional layers and convolutional layers to characterize the target,and then training the classifier for these layers separately.Finally,by searching the multi-layer response map,the target position is gradually located from coarse to fine.The algorithm is verified in the OTB-50 data set.A onetime evaluation(OPE)value can reach 0.612,which is better than the same algorithm.To extend the effectiveness of the algorithm on the local target of interest,the significance detection GBVS algorithm and the residual neural network ResNet-50 are fused.The single-layer feature layer tracking comparison experiment is designed and the effectiveness of the GBVS algorithm on the deep network ResNet-50 single-layer features is verified.The improved feature fusion tracking algorithm based on the significance detection GBVS algorithm is proposed.Algorithm is tested on OTB-50 data sets,which is basically superior to the same type of correlation filtering algorithm.Although the algorithm proposed in this paper has improved in accuracy and accuracy,there is still room for further improvement in real-time performance.The second part of the improved algorithm based on saliency detection increases the complexity of saliency detection and reduces the real-time performance of the algorithm,so the real-time performance needs to be further optimized in order to better balance the real-time and robustness of target tracking.