基于孪生卷积神经网络的目标跟踪算法研究

邹超 杨国平

摘 要： 传统的目标跟踪算法采用人工特征描述物体特征，这类人工设计的特征不能全面地表达一个物体的特点，在跟踪过程中这些特征点容易受到外界因素的影响，导致跟踪效果不稳定。基于卷积神经网络的目标跟踪算法由于采用卷积神经网络提取物体的深层次特征，这类特征能够模仿人脑描述学习一个物体的深层特征，使得在跟踪中具有较高的稳定性，目标不容易丢失且跟踪的准确性更高，能够适应复杂多变的环境鲁棒性更好。本文提出的算法采用Tensorflow搭建网络框架，离线训练模型，然后利用OpenCV调用训练好的模型进行目标跟踪实验。算法在确保较高的跟踪准确性基础上，又得到了较快的跟踪速度、且显示出很强的实时性，具有一定的实际应用价值。

关键词： 目标跟踪; 卷积神经网络; Tensorflow; OpenCV

文章编号： 2095-2163（2021）03-0177-04 中图分类号：TP183;TP391.41 文献标志码：J

【Abstract】The traditional target tracking algorithm uses artificial features to describe the characteristics of the object， this kind of artificial features can not fully express the characteristics of an object， so in the practical application of environmental changes these characteristic points will lead to inaccurate tracking， and the consequences become serious when the target is lost. The target tracking algorithm based on convolutional neural network uses convolutional neural network to extract the deep-level features of an object， which can describe the features of an object in a more comprehensive way， thus resulting in higher stability in the tracking process， less target loss and higher tracking accuracy， and being able to adapt to complex and changeable environments. The algorithm proposed in this paper uses Tensorflow to build network framework and offline train model， and then uses OpenCV to call the trained model for target tracking experiment. The algorithm not only keeps the tracking accuracy， but also has fast tracking speed and real time， which has certain practical application value.

【Key words】 target tracking; convolutional neural network; Tensorflow;  OpenCV

0 引 言

随着深度学习[1]的不断发展成熟，其应用范围也在不断扩大。最近很多学者将其引用到目标跟踪中。目标跟踪是计算机视觉领域的一个经典应用，通过跟踪单个或者多个移动目标，可以获取目标的位置、运动速度、方向等信息为相关的任务提供信息支持和消除误差。其应用诸如手势操作、行人行为预测、车辆跟踪、机器人抓手等。经典的目标跟踪算法主要有CamShift[2]为代表的运用目标的色彩特征的跟踪算法、以KCF[3]代表的核相关滤波算法、卡尔曼滤波算法[4]、粒子滤波算法[5]以及利用物体的特征点实现跟踪诸如基于SIFT特征点、SURF特征点的目标跟踪算法等。……