Research On Single Object Tracking Based On Deep Learning

A key area of computer vision is real-time Object Tracking,which is currently widely used in smart transportation and smart industrial technologies.Visual Object Tracking is divided into Single Object Tracking and Multi-Object Tracking according to the object quantity.In recent years,some breakthroughs have been made in the related algorithms of Single Object Tracking,but there are still many challenges such as scale change,occlusion,and boundary effects.There are still many theoretical issues that need further study.Thesis deeply analyzes the background of Single Object Tracking,improves related technologies and methods,and achieves phased results.Based on the Deep Learning algorithm Keep Track and the Siamese network structure,thesis introduces Ensemble Learning branch,Transformer-based encoder-decoder network and the untied positional encoding method,and uses the relevant dataset training to verify the tracking effectiveness.Compared with the classical tracking algorithm,the experimental results show that the improved Single Object Tracking accuracy is improved.The specific work content is as follows:First of all,in order to summarize context information in a larger spatial range and obtain better training results,when the model obtained by embedding Red Net-50 into the Keep Track algorithm extracts complex features,the involution operation in Red Net-50 will automatically Adaptive assignment to different locations prioritizes pheromones in the spatial domain.It brings great convenience to other subsequent modules of the tracker.Secondly,the cause of tracking failure in the tracking algorithm based on full convolution Siamese network in complex scenarios are analyzed,and the Ensemble Learning method is explored and designed.The PI control theory was introduced into Kmeans,and the uniformly distributed clustering samples were obtained.A clustering weight fusion module is designed,which adaptively assigns weights to each base tracker according to the semantic information of the target object,and fuses the base tracker into a strong tracker,which effectively solves the problem of tracking frame offset when the target occlusion occurs,and enhances the robustness and tracking accuracy of Siam FC.Finally,the Transformer feature fusion and the untied positional encoding strategy are applied to the Siam CAR algorithm.The encoder network splices and fuses the feature markers in the template image and the search image,and enhances the spliced feature markers layer by layer through the attention mechanism.The positional encoding helps the model to distinguish markers from different sources and different locations.feature map and feed it back to the prediction head network.Compared with SiamCAR,a more accurate bounding box estimation map is obtained.