| Most of the existing deep learning-based trackers use a global strategy to track the target,with the aim of learning a deep information representation of the entire object to locate it.However,tracking targets with various appearance changes is difficult for these methods.To solve this problem,a tracking strategy based on local information has emerged.This approach divides the target into multiple identical blocks and tracks all of them simultaneously,inferring the target state by combining the tracking results of these blocks.A potential limitation of this type of tracker is that not all blocks have the same tracking information,and some blocks may contain information that is not conducive to tracking.Besides,divide and conquer strategies are used to obtain information at different scales,but in the field of tracking,this strategy has serious performance problems.This article focuses on the aforementioned issues:Firstly,this article provides a detailed exposition on the research background and latest developments of current object tracking algorithms.It also introduces some fundamental concepts,such as neural networks,attention mechanisms,and dilated convolutions.Secondly,this article discusses the shortcomings of global information tracking strategies in existing target tracking algorithms,and analyzes the reasons why they are ineffective in dealing with factors such as target occlusion and appearance changes.To address this issue,a new cross-correlation method is proposed that leverages saliency feature mining for better target tracking and handling complex scenarios in target tracking,thus improving the performance of the model.Thirdly,this article analyzes the reasons for the poor accuracy of existing Siamese network-based object tracking algorithms and proposes a Hybrid Dilated Encoder(HDE)algorithm.The algorithm consists of two key components: the Hybrid Receptive Field Module(HRF)and the Hybrid Attention Module(HAM).The HRF module solves the performance problem of divide and conquer strategy in the tracking field by stacking the expansion convolution with different expansion rates,and makes better use of multi-scale information.The HAM module combines spatial self-attention and channel self-attention to better capture the internal correlation of features and reduce dependence on external information.Finally,this paper implemented the tracking algorithm model using the Py Torch framework and performed experimental validation and performance analysis on the Siamese network-based object tracking algorithm with saliency feature mining and the Hybrid Dilated Encoder algorithm using public datasets. |
