ThermiKit: Edge-Optimized LWIR Analytics with Agent-Driven Interactions
Lan Zeng, Chunhao Huang, Ruihan Xie, Zhuohan Huang, Yunqi Guo, Lixing He, Zhiyuan Xie, Guoliang Xing · 2025
Low-cost long-wave infrared (LWIR) cameras offer viable privacy-preserving perception for home environments, though deploying effective analytics on diverse, resource-limited devices remains challenging. We present ThermiKit, a plug-and-play thermal analytics stack with two key components: (1) an edge-optimized thermal sensing suite with a unified backbone and compact modules for detection, pose estimation, and tracking, adapted to common LWIR sensors through RGB→thermal transfer; and (2) a micro-MCP server that provides structured model outputs to large language models (LLMs) agents for natural-language interaction without transmitting raw images. Our evaluation uses both a new multi-camera parallel dataset and a long-term nursing-home deployment. The models deliver up to 49.34% mAP improvement over baseline performance while maintaining real-time operation on sub-1 TOPS devices, and the MCP-enabled agent correctly answers all logically formulated scene queries in our test set (10/10).