摘要
Abstract
Aiming at the problem of decreased vehicle tracking accuracy in occluded scenarios,this study proposes an algorithm named Convolutional Kernel Optimization for Occluded Vehicle Tracking(CKO-OVT).The CKO-OVT method employs a convolutional kernel optimization strategy to adaptively select convolution operators that are more sensitive to vehicle targets for feature extraction.It also utilizes a discriminative siamese network to evaluate tracking results and relocalize target when tracking fails,thereby further im-proving the robustness and accuracy of the tracking process.Experimental activities involve the construction of an Occluded Vehicle Tracking(OVT)database.Comparative experiments were conducted on public datasets Object Tracking Benchmark(OTB),TColor-128,and the self-built OVT dataset against nine leading algorithms:Efficient Convolution Operators for Tracking(ECO),Efficient Con-volution Operators for Tracking Using HOG and CN(ECOHC),Kernelized Correlation Filters Tracker(KCF),Discriminative Scale Space Tracker(DSST),Circulant Structure Kernel Tracker(CSK)Hierachical Convolutional Fectures for Visual Tracking(HCFT),Ro-bust Visual Tracking via Hierachical Convolutional Features(HCFTstar),Fully-Convolutional Siamese Networks for Object Tracking(SiameseFC),and Distractor-Aware Siamese Networks for Object Tracking(DaSiam)Experimental outcomes indicate that the CKO-OVT algorithm achieves improvements of 2.2%points in distance precision and 1.8%points in overlap success on the OTB dataset;0.4%and 0.9%percentage points in distance precision and overlap success,respectively,on the TColor-128 dataset;and 1.7%and 1.2%percentage points in distance precision and overlap success on the OVT dataset,respectively.The CKO-OVT proposed in this paper significantly improves the robustness and accuracy of vehicle tracking in occlusion scenarios through adaptive convolutional kernel optimization and a discriminative Siamese network.Experimental results on the OTB,TColor-128,and self-built OVT datasets demon-strate that the CKO-OVT algorithm outperforms mainstream tracking algorithms in both distance precision and overlap success rate,providing an effective solution for vehicle tracking in intelligent transportation and autonomous driving applications.关键词
卷积核优选/孪生网络/车辆跟踪/OVT数据集Key words
convolutional kernel/siamese network/vehicle tracking/OVT dataset分类
信息技术与安全科学
