To　address　the　challenges　in　E-learning　engagement　detection,　such　as　dependence　on　specialized　equipment,　intrusive　and　disruptive　methods,　and　limited　detection　dimensions,　a　novel　framework　based　on　behavioral　and　emotional　cues　for　E-learning　engagement　detection　is　proposed.　Utilizing　the　MediaPipe　module,　we　developed　four　feature　models:　head　posture,　blink　rate,　gaze,　and　facial　emotion.　Subsequently,　we　introduced　engagement　evaluation　metrics,　such　as　the　unit　time　distraction/smile　ratio,　and　captured　and　analyzed　learners’　behavioral　performance　and　emotional　changes　by　integrating　facial　feature　information　across　different　time　units.　Ultimately,　an　efficient,　low-cost,　and　easy-to-deploy　automatic　learner　engagement　recognition　system　was　created.　Experimental　results　demonstrated　that　the　proposed　head　posture　evaluation　model　outperformed　the　latest　machine　learning　methods　in　terms　of　accuracy,　robustness,　and　computational　efficiency.　The　blinking　detection　model　based　on　dynamic　EAR　thresholds　improved　the　limitations　of　the　current　fixed　EAR　threshold　method　in　handling　individual　differences.　The　gaze　tracking　model　exhibited　superior　overall　performance　compared　to　the　latest　computer　vision　schemes　based　on　the　MediaPipe　iris　module.　The　smile　detection　model　based　on　XGBoost　and　MediaPipe　facial　landmarks　also　outperformed　deep　learning　methods　in　accuracy,　training,　and　inference　time.　Overall,　compared　to　other　recent　research　achievements,　the　proposed　models　achieved　satisfactory　recognition　accuracy,　higher　computational　efficiency,　and　lower　hardware　requirements,　making　them　suitable　for　widespread　deployment　on　edge　devices　and　demonstrating　their　practical　application　value　in　E-learning　scenarios.　©　2025　Elsevier　Ltd