| Single object tracking is one of the key tasks in the field of computer vision,which is widely used in intelligent video surveillance,military security,autonomous driving,intelligent human-computer interaction,ocean exploration and other fields.Therefore,the research on visual object tracking has important practical significance.Currently,in the application of single object tracking in public area scenes,achieving long-term and stable object tracking still faces great challenges caused by complex factors such as occlusion,background clutter,and scale variation.This thesis focuses on the problem of single object tracking under background clutter,scale variation,and other situations.Specifically,a single object tracking algorithm based on deep learning is proposed.The specific works are as follows:(1)Aiming at the problems of background clutter and occlusion in single object tracking in public scenes,an Anchor-Free single object tracking algorithm based on channel attention is proposed.First,the proposed Parallel Pooling Channel Attention(PPCA)module is embedded in the template and search branch,and the channel weights of template features and search region features are rescaled by using the relationship among the feature channels.In this way,the network’s ability to discriminate foreground and background can be further improved.Second,a decentralized mapping strategy is proposed.According to the different regions where the object centers point are located,different center offsets are set to map the object center points on the classification feature maps to the search regions more accurately,so as to locate the object more accurately.Expensive experiments on the UAV123,OTB100 and GOT-10 k datasets show that the proposed algorithm can effectively alleviate the probles of background clutter and occlusion.(2)Aiming at the problems of object scale variation and deformation in natural scene single object tracking,a graph attention single object tracking algorithm based on Poly Loss is proposed.First of all,in order to extract more robust features,the algorithm embeds the proposed plug-and-play PPCA module in the last layer of the backbone network.With the graph attention information transmission framework,it can better deal with the problems of object deformation and scale variation.Secondly,a flexible loss function framework Poly Loss is introduced.By fine-tuning a polynomial base,the loss function is more suitable for single object tracking tasks and the accuracy of the algorithm is improved.Experiments on the VOT2016 and GOT-10 k datasets show that the proposed algorithm can effectively overcome the obstacles of target scale variations and deformations.(3)Based on the PyQt5 framework,a single object tracking application system is designed and implemented.According to different tracking requirements,a variety of tracking algorithms are embedded in the system.Model switching,offline/online tracking mode selection,save video back to watch and other functions are realized.Finally,functional and non-functional tests are carried out on the system to verify the stability,practicability and ease of operation of the developed system. |
