基于多特征提取和自适应关键滤波器的目标跟踪算法  

Object Tracking Algorithm Based on Multi-feature Extraction and Adaptive Key Filter

作　　者：刘惠临[1] 轩文杰 LIU Huilin;XUAN Wenjie(School of Computer Science and Engineering,Anhui University of Science and Technology,Huainan 232001,China)

机构地区：[1]安徽理工大学计算机科学与工程学院,安徽淮南232001

出　　处：《湖北民族大学学报（自然科学版）》2023年第4期462-470,共9页Journal of Hubei Minzu University:Natural Science Edition

基　　金：安徽省高等学校自然科学研究重点项目(KJ2020A0299)。

摘　　要：为解决核相关滤波(kernelized correlation filter,KCF)目标跟踪算法在复杂场景下精准度和成功率低的问题,提出基于多特征提取和自适应关键滤波器的目标跟踪算法(multi-feature extraction and adaptive key filter based object tracking algorithm,MEAKF)。首先,选择帧图像的方向梯度直方图(histogram of oriented gradients,HOG)特征、深度浅层卷积特征以及深度高层卷积特征,通过主成分分析(principal component analysis,PCA)对特征进行降维,将这3种特征分别训练得到相应的滤波器。然后,基于跟踪目标感兴趣区域的相似度对视频序列中的关键帧图像进行判断,关键帧图像上采用深度高层卷积特征训练的滤波器作为关键滤波器。最后,在非关键帧图像上采用2种滤波器进行目标定位,在关键帧图像上采用3种滤波器进行目标定位,以此达到更好的跟踪效果。MEAKF算法与KCF算法相比,在精准度上提升了0.051,在成功率上提升了0.104,算法性能得到了有效提高。To address the problem that the kernelized correlation filter(KCF)object tracking algorithm is low in terms of accuracy and success rate in complex scenes,we propose multi-feature extraction and adaptive key filter based object tracking algorithm(MEAKF).Firstly,choose the directional histogram of gradients(HOG)features,deep shallow convolutional features and deep high-level convolutional features of the frame image,reduce the dimensionality of the features by principal component analysis(PCA),and train each of the three features to get the filters.Then,judge the key frame images in the video sequence based on the similarity of the tracking object region of interest,and use the filters trained with the deep high-level convolutional features as the key filters on the keyframe images.Finally,use two filters for object localization on non-keyframe images,and three filters for object localization on keyframe images,thus achieving better tracking results.MEAKF algorithm improves 0.051 in accuracy and 0.104 in success rate over KCF,effectively improving algorithm performance.

关 键 词：机器视觉 目标跟踪 滤波 特征融合 自适应关键帧 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]