Method Of Siamese Network Based On Template Online Update And Enhanced Cross Attention For Object Tracking

Target tracking is one of the fundamental research topics in the field of computer vision.This task aims to accurately locate a target of interest in a video sequence and evaluate its scale with a rectangular frame.This technique is widely used in the fields of video surveillance,unmanned driving,and machine interaction.Most of the current tracking algorithms use Siamese network to handle the target tracking task for its great robustness and accuracy.However,the discriminative model-based Siamese network still has two main problems:1)When the target is occluded,the tracker will continuously track the occluded object as the correct target,thus losing the tracked target;2)The feature extraction network of traditional methods cannot fully utilize the contextual information,which tends to reduce the success rate and accuracy of tracking in complex scenes.To address the above problems,this paper conducts the following studies:(1)To solve the problem of target loss in the actual tracking processing of the tracker,this paper proposes to add an online update module based on key frames to the tracking algorithm.Different from the traditional online update mechanism,the algorithm selectively extracts key frames as new templates in the tracking process to participate in the similarity calculation.The structure mainly includes:a threshold-based keyframe discriminator;a global search module;an optimizer integrated with online training and iteration.(2)To solve the problem that Siamese network cannot utilize contextual relationships,this paper proposes to add Transformer-based encoder-decoder modules to the Siamese network.The module is designed in a parallel manner to perform cross attention on template features and search features using Multi-Head Attention,and output more discriminative cross-features that are used for classification and regression.In this paper,several modules are proposed to be applied to the basic Siamese network,mainly including:superposition fusion module;attention head network;feature enhancement module.In the experiments,ResNet-50 and AlexNet are selected as the main networks,and the CoCo,ILSVRC2015,LaSOT,and GOT-10k datasets are used for training to obtain pre-trained models,which are tested on several publicly tracking test sets(OTB100,VOT2018,etc.).The experimental results show that the online update module proposed in this paper can improve the robustness of the algorithm and reduce the probability of target loss;and the method based on enhanced cross attention effectively improves the overall performance of the algorithm,which can deal with the tracking tasks in most scenarios.