基于卷积神经网络的实时跟踪算法

Real-time tracking algorithm based on CNN

利用卷积神经网络(CNN)强大的特征学习能力,提出了一种基于卷积神经网络的实时跟踪算法。通过对双通道卷积神经网络进行离线训练,学习相邻两帧之间的差异,得到跟踪目标的表观特征与运动之间的普遍规律。在不需要对网络模型在线更新的情况下,直接通过网络回归得到对目标的位置和对应置信度的预测。在VOT2014数据集中进行实验,结果表明:提出的跟踪算法的性能达到了当前领先水平。同时,跟踪算法的运行速度可以达到90帧/s,表现出非常不错的实时性。

A real-time tracking algorithm based on convolutional neural network（CNN） is proposed,using strong capabilities of feature learning of CNN. In this algorithm, a generial rule between motion and appearance features of tracking target is obtained, by offline training on 2-channel CNN , difference of two adjacent frames is learned, directly by network regresses obtain prediction of location of the target and the confidence score. In case of no online update is required. Results of experiment on VOT2014 dataset show that performance of the proposed tracking algorithm achieves the state-of-the-art, at the same time,it operates at a very high speed of 90 frames-per- second,which show good realtime performance.