○ 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”.

○ 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.

○ 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.