BROWSE
Detail
Conference Paper
Natural Language-Driven Active AI Agent Camera System: LLM-Guided Semantic Anchor Modulation for Real-Time Edge Deployment
Cited - time in 
Share 
Share
- Authors
- Hyan su Bae, Jinhong Kim, Yun-Won Choi, Jang Woon Baek
- Issue Date
- 2026-07
- Citation
- International Conference on Ubiquitous and Future Networks (ICUFN) 2026, pp.1-6
- Publisher
- IEEE
- Language
- English
- Type
- Conference Paper
- Abstract
- This paper proposes an active AI agent camera system that interprets natural-language queries in real time on a CCTV-attached edge device, detects and tracks a specified target, and autonomously generates camera motion commands. The core contribution is LLM-Guided Semantic Anchor Modulation (LGSAM), a mechanism that injects semantic attribute vectors parsed by a large language model and encoding color, clothing, gender, and direction into the detector's anchor feature maps via channel attention, enabling lightweight open-vocabulary detection. LGSAM achieves a 4.6× inference speedup on Nvidia Orin AGX relative to Grounding DINO while limiting accuracy loss to under 2.4 percentage points and sustaining a practical 26.7 FPS on Orin Nano. Five algorithm combinations are benchmarked on both the Nvidia Orin AGX 64GB and Orin Nano 8GB platforms, and the full system is validated through two natural-language-guided person-tracking experiments on real indoor CCTV footage.
- Keyword
- Active AI agent camera, LLM intent parsing, Semantic anchor modulation, Open-vocabulary detection, Multi- object tracking, Camera command generation
- KSP Keywords
- AI Agent, Accuracy loss, Camera system, Command generation, Edge devices, Feature Map, Language Models, Natural language, Object Tracking, Percentage points, Person tracking