As industrial automation progresses, collaborative human-robot interaction (cHRI) is becoming more prevalent, making it essential to understand the factors influencing these interactions. One key factor is physical human fatigue, which reduces muscular force capacity and impacts both productivity and safety. While recent research emphasizes wearable sensors for direct measurement, non-intrusive camera-based approaches remain underexplored, despite their potential for more natural fatigue tracking. This paper examines visual cues, such as body posture and facial expressions, to assess their effectiveness in classifying fatigue states in cHRI. We demonstrate that a simple long short-term memory (LSTM) network performs well, particularly with posture data, and that facial features provide further insights into fatigue onset. Our findings suggest a social dimension, where robots that recognize and respond to human fatigue could enhance collaboration in industrial environments.