Research On Meta Transfer Learning Based Object Tracking Algorithm

It enables the computer to stably process the video sequence captured by the camera in real time,accurately locate the target coordinates and draw the appearance boundary,which is the core requirement of the visual tracking task.Different from laboratory research,the industry has more stringent standards for the balance of real-time performance and accuracy of trackers,as well as generalization ability.In recent years,traditional tracking models usually sacrifice real-time performance in order to obtain considerable accuracy improvements.In contrast,the Siamese network tracker can effectively balance accuracy and speed,and provide a reference solution for the industrial expansion of tracking methods.Considering the superiority of the Siamese network,object tracking based on the Siamese network has maintained a high research interest for many years.This thesis focuses on the tracking method based on the Siamese network.In view of the inherent redundant parameters of the fully convolutional tracker SiamFC,which leads to overfitting and low generalization ability of general targets,corresponding solutions are designed respectively.The main points and achievements of the research are as follows:(1)Aiming at the problem of over-fitting caused by redundant parameters of the backbone network of SiamFC tracker,redundant features are pruned according to the attention mechanism in the backbone network.Through systematic analysis,it is found that spatial attention can achieve effective network pruning in visual tasks trained by labeling,but in the SiamFC tracker based on regression training,the spatial attention mechanism will cause the problem of distraction of the backbone network.Channel attention aims to establish the associativity of the backbone network channels,thus enabling focused attention in regression problems.Based on this,the subject focuses on the network pruning strategy based on channel attention.Simulation experiments show that for the backbone network Alex Net used to keep the tracker model lightweight,the spatial attention sub-CBAM will degrade the performance of the tracker on the OTB-100 dataset,while the channel attention SE-Block has a significant impact on the success rate and accuracy rate,by 1.9% and3.7%,respectively.Therefore,channel attention can prune redundant parameters and efficiently release the channel feature expression potential of the backbone network.(2)In view of the low generalization ability of the SiamFC tracker in new scenarios,the idea of transfer learning is introduced into the tracking task.In the meta-learning stage,only the meta-parameters that control the basic neurons are updated,making the meta-update operation lightweight,and meta-learning can only rely on a limited number of video sequences to complete the generalization of the tracker.The simulation results show that the success rate of the MTL-based SiamFC tracker on the OTB-100 dataset is 1.5% higher than that of(1),and the accuracy is 1.9%higher.Therefore,meta-learning MTL has a positive effect on the efficient generalization of the tracker to new scenarios.