School of IoT Engineering, Jiangnan University, Wuxi, Jiangsu
随着目标跟踪技术在多种视觉任务中的广泛应用，跟踪算法的实时性变得越来越重要。全卷积孪生网络跟踪算法（SiamFC）虽然在跟踪速度方面较为理想，但是在应对复杂的跟踪环境时很容易出现跟踪漂移，为了能在提高算法精度的同时保证实时性，提出一种基于负样本挖掘与特征融合的高速跟踪算法。首先，为了学到更深层次特征的同时，不过多增加额外参数运算，使用改进后的ShuffleNetV2轻量级网络进行特征提取，提升跟踪速度；其次，在离线训练阶段引入不同种类的负样本对，加强对语义信息的学习，从而提升模型的特征判别能力；最后，为了得到更高质量的响应图，提出一种多尺度特征融合策略，充分利用浅层与深层特征，提高跟踪精度。在OTB100和VOT2018两个数据集上与其他跟踪算法进行对比实验，结果表明，所提算法较基准算法SiamFC在各项指标上有大幅度提升，在两个数据集下分别收获了8.3%和7.9%的增益。同时在NIVIDA GTX l070下的速度可达114FPS。
With the wide application of object tracking technology in a variety of vision tasks, the real-time requirements for tracking algorithms are becoming increasingly important. Although the fully-convolutional siamese network algorithm for object tracking (SiamFC) is ideal in tracking speed, it is prone to tracking drift when dealing with complex tracking environment. In order to simultaneously improve the tracking accuracy and speed, a high-speed tracking algorithm based on negative example mining and feature fusion is proposed. Firstly, the improved ShuffleNetV2 lightweight network is used for feature extraction, which can learn deeper features without adding extra parameters and calculations. Secondly, different types of negative example pairs are introduced in the offline training phase to strengthen the learning of semantic information, aiming at improving the feature discrimination ability of the model; finally, a multi-scale feature fusion strategy is adopted, which make full use of shallow and deep features to obtain a higher quality response map and greatly improve tracking accuracy. The experiment results of OTB100 and VOT2018 datasets show that the proposed algorithm significantly outperforms the benchmark algorithm SiamFC in various indicators, yielding 8.3% gain in OTB100 dataset and 7.9% gain in VOT2018 dataset. At the same time, the proposed tracker can perform 114FPS under NIVIDA GTX 1070.