结合双线性特征融合与自适应重检测的目标跟踪方法  

Object Tracking Method Combining Bilinear Feature Fusion with Adaptive Re-Detection

作　　者：闫河 张唯 刘宇涵 黄奎霖 李尧 Yan He;Zhang Wei;Liu Yuhan;Huang Kuilin;Li Yao(Liangjiang College of Artificial Intelligence,Chongqing University of Technology,Chongqing 401135)

机构地区：[1]重庆理工大学两江人工智能学院,重庆401135

出　　处：《计算机辅助设计与图形学学报》2024年第11期1691-1699,共9页Journal of Computer-Aided Design & Computer Graphics

基　　金：国家重点研发计划(2018YFB1308602);国家自然科学基金(61173184);重庆市自然科学基金(cstc2018jcyjAX0694);重庆市研究生科研创新项目(CYS23694)。

摘　　要：针对SiamRPN目标跟踪方法用一阶浅层网络提取特征,难以精确地获得丰富的特征信息;缺少遮挡判别机制,易导致目标漂移或跟丢的问题,提出一种双线性特征融合与自适应重检测相结合的孪生网络目标跟踪方法.使用改进的ResNet50提取序列特征,对最后3个残差块提取的特征进行双线性级联融合,获得二阶特征信息,并通过区域候选网络输出目标框;计算目标框对应的平均峰值相关能量,判断目标是否被遮挡;针对遮挡,构建以上一帧跟踪结果为中心的邻近检测窗口,结合权重顺序选择与随机选择的方式选取窗口,对目标重检测.对比OTB100和UAV123数据集上的实验结果表明,所提方法跟踪成功率分别达到0.894和0.800,跟踪精确度分别达到0.669和0.605,同时具有较好的跟踪时效性.The SiamRPN tracker,relying on a first-order shallow network for feature extraction,encounters challenges in accurately capturing comprehensive feature information,often resulting in target drift or loss due to the absence of an occlusion discrimination mechanism.To address these issues,this paper introduces a twin-network tracking method that combines bilinear feature fusion with adaptive re-detection.We em-ployed an enhanced ResNet50 for sequential feature extraction,with feature vectors obtained from the final three residual blocks fused in a bilinear cascade,thereby providing second-order feature information.Sub-sequently,the region proposal network generates the target box.To evaluate potential occlusion,we calculate the average peak correlation energy corresponding to the target box.In cases of occlusion,a neighboring detection window is established around the tracking result from the previous frame.The window’s selection is determined through a combination of weighted sequential and random selection for target re-detection.Experimental results on the OTB100 and UAV123 datasets demonstrate the effectiveness of our proposed method,achieving tracking success rates of 0.894 and 0.800,as well as tracking accuracies of 0.669 and 0.605,respectively.Furthermore,the method exhibits robust tracking timeliness.

关 键 词：目标跟踪 孪生网络 抗遮挡 重检测 特征融合 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]