无锚双注意力孪生网络的视觉跟踪

针对跟踪过程中因光照变化、快速运动及尺度变化等造成的角点定位精准度下降问题,受SiamCAR的跟踪框架启发提出一种无锚双注意力孪生网络的视觉跟踪算法.首先,算法的主干网络采用ResNet-50并结合增强多层融合特征图进行特征提取,充分利用网络浅层特征的定位信息和深层次的语义信息,提高算法对目标特征的语义理解能力;然后,构建混合注意力模块缓解无锚跟踪器角点定位不准确问题,提高算法的跟踪准确性和定位精度;最后,在GOT10K、UAV123、LaSOT等数据集上进行广泛实验,并与当前的先进跟踪器进行比较,该算法可以较好地抵抗光照变化、快速运动及尺度变化等多种复杂因素带来的影响,同时,在多项评测指标上获得了良好的跟踪性能.

Dual attention Siamese network with anchor free for visual tracking

Aiming at the problem of decreased corner positioning accuracy caused by light changes, fast motion and scale changes in the tracking process, we propose a visual tracking algorithm motivated by the framework of SiamCAR. Firstly, the research method uses improved ResNet-50 as a feature extraction backbone network and combines with enhanced multi-layer fusion feature map to extract feature, which makes full use of the location information of shallow features and deep semantic information of the network, and improves the semantic understanding ability of the algorithm to target features. Secondly, a hybrid attention block is constructed to alleviate the inaccurate corner location of the anchor tracker, which improves the tracking accuracy and positioning accuracy of the algorithm. Finally, extensive experiments are carried out on GOT10K, UAV123, LaSOT and other datasets. Besides, compared with current advanced trackers, the proposed algorithm can better resist the influence of various complex factors such as illumination variation, rapid motion and scale variation, at the same time, obtain good tracking performance on a number of evaluation indicators.