首页|期刊导航|现代电影技术|融合改进MobileFaceNet与时序关联的影视人物身份追踪算法

融合改进MobileFaceNet与时序关联的影视人物身份追踪算法OA

A film and television character identity tracking algorithm integrating and im-proving MobileFaceNet with time-series correlation

中文摘要英文摘要

随着影视制作领域大量非结构化视频数据的激增,如何从复杂场景中高效、准确提取特定角色的镜头素材,已成为影视后期处理场景的一个难题.传统识别方案常在特效妆造掩盖、光照剧变以及蒙太奇式的频繁剪辑时失效而导致人物轨迹断裂.本研究构建了一套融合改进MobileFaceNet与时序关联逻辑的人物身份追踪算法,通过向轻量级MobileFaceNet网络中嵌入卷积块注意力机制,增强模型在复杂妆造下的特征判别力,同时改进了一种融合运动预测与表观特征相似性的级联匹配策略,并对蒙太奇剪辑进行分析设计了基于全局特征库的身份重识别方法,有效解决了角色轨迹断裂问题.算法在自建影视视频数据集上的实验数据显示,其多目标追踪准确度达到 78.9%,身份综合得分达到78.8%.该模型极大地减少了身份频繁切换的错误,同时兼顾了极快的模型推理速度,符合影视工业的素材索引高精度要求与实时性需求.

With the surge in unstructured video data within film and television production,efficiently and accurately extract-ing footage of specific characters from complex scenes has become a major challenge in post-production.Traditional identi-fication methods often fail when confronted with special effects makeup,drastic lighting changes,and frequent montage-style editing,resulting in fragmented character trajectories.This study proposes a character identity tracking algorithm that integrates an improved MobileFaceNet with temporal correlation logic.By embedding a convolutional block attention mechanism into a lightweight MobileFaceNet network,the model's feature discrimination capability is enhanced under com-plex makeup conditions.Simultaneously,the study improves a cascaded matching strategy that incorporates motion predic-tion and appearance feature similarity,and analyzes montage editing to design an identity re-identification method based on a global feature library,effectively addressing the problem of trajectory fragmentation.Experiments on a self-built film and television video dataset show that the proposed method achieves a multi-target tracking accuracy of 78.9%and an overall identity score of 78.8%.Notably,the model significantly reduces errors caused by frequent identity switching while maintaining a fast inference speed,thereby meeting the high-precision and real-time requirements of the film and television industry for material indexing.

徐敏;张珍奇;夏天

上海第二工业大学计算机与信息工程学院,上海 201209上海第二工业大学计算机与信息工程学院,上海 201209上海第二工业大学计算机与信息工程学院,上海 201209

社会科学

影视人物追踪MobileFaceNet注意力机制身份重识别

Film and Television Character TrackingMobileFaceNetAttention MechanismIdentity Re-Identification

《现代电影技术》 2026 (4)

55-63,9

上海市教委青年教师资助计划"《大模型应用技术专业导论》课程思政探索与实践研究"(A01GY25F012).

10.3969/j.issn.1673-3215.2026.04.008

评论