The long-term accumulation of mental stress significantly impacts an individual’s physical and mental health as well as workplace safety. Identifying mental stress naturally and non-intrusively in daily work and life holds substantial research significance and practical value. Addressing the limitations of traditional mental stress identification methods—which rely on subjective questionnaires or contact-based physiological sensors and struggle to meet non-intrusive application requirements—this paper proposes an auxiliary mental stress identification method based on implicit information from facial videos. This method utilizes remote photoplethysmography (rPPG) as its foundation. It extracts subtle color variation signals caused by skin reflection from facial video sequences to reconstruct rPPG signals and extract key physiological features. Additionally, it incorporates facial action unit (AU) features through a dual-branch feature fusion approach to achieve mental stress recognition. Furthermore, a facial video dataset under various stress-inducing conditions was constructed based on the Montreal imaging stress task (MIST) paradigm to validate the proposed method’s effectiveness. Experimental results demonstrate that the proposed method achieves a mental stress recognition accuracy of 95.64%. Furthermore, the findings reveal consistent patterns of change in pulse wave modulation and facial movement expression corresponding to mental stress states at the feature level, providing an effective technical approach for non-intrusive and interpretable mental stress recognition.