○ Project goal: Build an intelligent-agent closed loop based on React-Loop, reconstruct video object tracking into an Agentic AI process that is “plannable, verifiable, and self-checking”, lower the barrier for non-experts in category selection, threshold tuning and ID assignment, and realize an end-to-end experience of “natural-language target specification → automatic tracking → automatic editing output”.○项目目标：构建基于React-Loop的智能体闭环，将视频目标追踪重构为“可规划、可验证、可自校验”的Agentic AI过程，降低非专家在类别选择、阈值调参与ID分配上的门槛，实现“自然语言指定目标→自动追踪→自动剪辑输出”的端到端体验。

○ System/algorithm design: Build a cascaded system of “instruction parsing → existence verification → detection-tracking-reverification closed loop → trajectory bridging and editing”: use DeepSeek to parse natural language to generate YOLO category sets and CLIP prompts; introduce Gemini-2.5 Flash for target existence verification and initial detection parameter recommendation; the execution layer adopts YOLO multi-scale detection + ByteTrack identity preservation (scene-change reset + CLIP semantic filtering and cross-segment global ID merging), and exports temporally coherent video clips aligned with intent through trajectory bridging.○系统/算法设计：构建“指令解析→存在性验证→检测-追踪-重验证闭环→轨迹桥接与剪辑”级联系统：使用 DeepSeek解析自然语言生成YOLO类别集合与CLIP提示；引入Gemini-2.5 Flash进行目标存在性核验并推荐初始检测参数；执行层采用YOLO多尺度检测+ByteTrack身份保持（场景变化reset+CLIP语义过滤与跨片段全局ID合并），并通过轨迹桥接导出与意图对齐的时序连贯视频片段。

○ Effects and contribution: Achieved average Recall 75.6% / Precision 92.7% / Coverage 92.5% on 14 test cases; verified the feasibility of the paradigm of “semantic-guided tracking + closed-loop reverification + automatic editing”, and identified the main bottlenecks of multimodal matching in scenarios with overly strong attribute descriptions / semantic ambiguity.○效果与贡献：在 14 个测试用例上取得平均召回率Recall 75.6% / 精准率Precision 92.7% / 覆盖率Coverage 92.5%；验证“语义引导追踪+闭环重验证+自动剪辑”范式的可行性，并识别多模态匹配在属性描述过强/语义歧义场景下的主要瓶颈。