自适应博弈融合的记忆修复跟踪算法研究

Research on Memory Repair Tracking Algorithm with Adaptive Game Fusion

针对跟踪任务中出现的目标遮挡和短暂消失等挑战属性,提出了一种结合记忆修复机制的特征博弈融合目标跟踪算法。首先,在核相关滤波器的框架下,提取目标的FHOG特征、SURF特征进行通道串联,再自适应博弈融合ResNet-50网络提取的Conv4-1层深度特征共同构建目标外观模型,避免跟踪算法的特征单一性。然后,改进模型更新策略,通过记忆修复神经元和记忆矩阵实现对特征信息的记忆、回忆和修复,完成遮挡或漂移视频帧的特征损失修复工作,使用修复完整的特征模板对核相关滤波器进行正确更新,提高算法对跟踪对象的辨识度。最后,在OTB100、TC128、UAV123实验基准数据集上与KCF、DCF_CA、SAMF、M2C2F和EMCF算法进行对比测试,实验结果表明改进算法的跟踪准确性和鲁棒性得到提高。

Aiming at the challenging attributes such as target oeclusion and transient disappearance intracking tasks,a feature game fusion target tracking algorithm combined with memory repair mechanism isproposed.Firstly,under the framework of kemel correlation filter,the FHOG,features and SURF features ofthe target are extracted for channel series,and the Conv4-1 layer depth features extracted by ResNet-50 arefused by adaptive game to jointly construct the target appearance model to avoid the feature singularity of thetracking algorithm.Then,the model update strategy is improved to memorize,recall and repair the featureinformation through the memory repair neurons and the memory matrix,and the feature loss repair work of theoccluded or drifting,video frames is completed,The repaired feature template is used to correctly updatethe kernel correlation filter to improve the recognition of the tracking object.Finally,the proposed algorithmis compared with KCF,DCF _CA,SAMF,M2C2F and EMCF algorthms on 0TB100,TC128 and UAV123 experimental benchmark datasets.The experimental results show that the tracking accuracy and robustness of the improved algorithm are improved.