This study proposes the development of a driver monitoring and interactive AI agent system on edge devices. The system utilizes lightweight large language models for decision-making and operations, as well as a driving behavior recognition system to monitor driving safety. Among these, the ADAS Buddy intelligent vehicle agent system provides an intuitive human-machine interaction interface, while TSM-MobileNetV4 effectively identifies distracted driving and mobile phone usage behaviors. The Multi-Task Model (DMTM) can estimate eye state, head posture, and gaze direction simultaneously, providing comprehensive monitoring and guidance for driver behavior safety.

By implementing lightweight language models with 4 billion and 8 billion parameters, this study developed a highly specialized, high-performance automotive AI agent with hardware requirements that are less than 1% of those of GPT-4o. This enables drivers to minimize distractions and complete tasks autonomously. The Temporal Shift Module (TSM), when used in conjunction with a lightweight architecture, delivers high-performance driving behavior recognition. The DMTM model integrates multiple subtasks, reducing computational complexity and model parameters without sacrificing accuracy, and outperforming existing methods.

This system offers the following advantages: high performance, low computational load, and minimal parameters. These features make it particularly suitable for deployment on embedded platforms. It can maintain complete service stability even when network services are unstable, effectively enhancing driving safety. This technology serves a dual purpose: it can be used as a driver assistance system in vehicles and as a tool for fleet management and the logistics industry. It can monitor driver status and prevent accidents.