摘要
为提高目标形变、遮挡、相似干扰以及视野超出等复杂场景的目标跟踪性能,提出一种复杂场景单目标跟踪算法。基于Staple算法研究二维高斯函数像素权重赋予问题,优化颜色直方图统计,增强目标与背景区分度。引入基于峰值旁瓣比(Peak Side Lobe Ratio,PSR)的HOG特征、颜色特征的自适应融合机制,合理选择融合系数,确保混合特征更加可靠。分析目标区域中心与上一帧目标中心距离,结合最大混合响应计算最佳中心位置,解决相似目标干扰问题。采用混合响应、HOG特征、平均峰值相关能量(Average Peak-to-Correlation Energy,APCE)判定目标丢失、遮挡情况,保持目标框位置,实现目标的及时重新跟踪。采用结合之前帧和当前帧信息的模板更新策略,进一步提升跟踪精度,并在OTB100数据集中涉及形变、遮挡、视野超出3个属性视频上测试。实验结果表明,改进算法在整体和特定属性(形变、遮挡、出视野)的成功率及形变属性的精确度上,较Staple算法分别提升了1.8%,3.3%,2%和9%;在VOT16数据集上,改进算法在整体和遮挡属性上,重叠度较Staple提升了0.0222和0.0196,满足复杂的特定场景下的目标跟踪需求。
To address challenges in single-target tracking under complex scenarios such as target deformation,occlusion,similar interference,and out-of-view situations,a novel tracking algorithm is proposed.Building on the Staple algorithm,the method optimizes pixel weight assignment using a two-dimensional Gaussian function and enhances the color histogram to improve target-background distinguishability.An adaptive fusion mechanism based on the Peak Side Lobe Ratio(PSR)is introduced to combine HOG and color features,with carefully selected fusion coefficients ensuring feature reliability.The target's optimal center position is determined by analyzing the distance between the current and previous frame centers, alongside the maximum composite response, effectively mitigating interference from similar targets. Tar get loss or occlusion is identified using composite response, HOG features, and Average Peak-to-Correla tion Energy (APCE), maintaining the target frame's position and enabling timely re-tracking upon reap pearance. A template update strategy combining past and current frame information further enhances track ing accuracy. Tests on the OTB100 dataset with deformation, occlusion, and out-of-view scenarios show that the improved algorithm increases overall and specific attribute success rates (deformation, occlusion, out-of-view) by 1. 8%, 3. 3%, 2%, and deformation precision by 9% compared to the Staple algorithm. On the VOT16 dataset, the overlap rate for overall and occlusion attributes improves by 0. 022 2 and 0. 019 6 respectively, meeting the demands of target tracking in complex scenarios.
作者
林慧兰
赵春蕾
郝志成
刘时
朱明
姜鑫
高文
张军强
LIN Huilan;ZHAO Chunlei;HAO Zhicheng;LIU Shi;ZHU Ming;JIANG Xin;GAO Wen;ZHANG Junqiang(Changchun Institute of Optics,Fine Mechanics and Physics,Chinese Academy of Sciences,Changchun 130000,China;BYD Auto Industry Company Limited,Shenzhen 518000,China)
出处
《光学精密工程》
CSCD
北大核心
2024年第23期3490-3503,共14页
Optics and Precision Engineering
基金
国家重点研发计划资助项目(No.314)
吉林省科技发展计划重点研发项目(No.20220201146GX)。
关键词
单目标跟踪
复杂场景
背景抑制
相似目标再识别
丢失判定
single object tracking
complex scenarios
background suppression
similar target re-identification
loss determination
