稀疏卷积特征的实时目标跟踪

Sparse Convolutional Features for Real-time Object Tracking

针对分层卷积相关滤波目标跟踪算法鲁棒性良好而实时性能较差问题,提出了基于单层稀疏卷积特征的实时目标跟踪算法.该方法选用单个卷积层的特征,并通过等间隔采样方法生成稀疏特征来提高跟踪的速度.然后采用调整类标函数带宽的策略来提高核相关滤波分类器性能,以解决卷积特征维度降低造成的算法鲁棒性下降问题,在OTB-2013标准数据集上测试算法.实验结果表明,该算法的平均距离精度为89.9%,平均跟踪速度为25.0帧/秒,比原分层卷积核相关滤波目标跟踪算法分别提高了0.9%和108.3%;与多尺度域适应的目标跟踪算法速度相比稍有降低,但平均距离精度提高3.8%,在目标发生遮挡、形变、光照变化、背景混乱等情况时具有很好的鲁棒性.

A real-time object tracking algorithm based on single-layered convolutional features is proposed to improve the speed of hierarchical convolutional feature tracking. The new algorithm first selects single-layered convolutional features and then takes equal interval sampling method to generate sparse features,which lead to a considerable improvement on the tracking speed. Then the kernel width of Gaussian label function is adjusted to improve the performance of the correlation filter classifiter in order to solve its poor robustness caused by dimension reduction of convolutional features. Extensive experimental results on OTB-2013 benchmark dataset show that the mean distance precision of the proposed algorithm is 89. 9% and the speed is 25 frames per second,outperforming hierarchical convolutional features tracking by 0. 9% and108. 3% respectively. Compared with object tracking via multi-scale domain adaptation,the mean distance precision increased by 3. 8%,though its speed declines slightly. In the case of occlusion,deformation,illumination variation,background clusters,etc.,this new algorithm executes robust performances.