Real-time monitoring of student behavior in classrooms aids in dynamically assessing learning status and optimizing the learning experience. However, this remains challenging in crowded and dynamic classroom settings where small targets are prevalent. To confront these obstacles, we introduce YOLO-RTCM, a refined and real-time capable detection model based on YOLOv12, which is particularly optimized for monitoring student activities in dynamic educational settings. The model enhances feature extraction in complex backgrounds through the Context Guided approach, resolves multi-scale occlusion problems using MultiSEAM, and improves generalization and detection accuracy through the use of Inner-CIoU. Furthermore, we have constructed a diverse student classroom behavior dataset that covers various complex classroom scenarios, with each student manually annotated using bounding boxes and behavior classifications, providing a reliable benchmark for training and evaluation. Experimental results demonstrate that YOLO-RTCM outperforms benchmarks with 0.766 mAP @0.5 and balanced metrics (200BF1: 0.722, Recall: 0.720), successfully addressing challenges of environmental variance and significant student size variations in complex classrooms while ensuring real-time efficiency.