Nowadays, the number of applications of IoT and AI has increased rapidly to provide personalized learning environments that provide control to learners. To promote good sitting posture and support the health of primary and middle school students, intelligent learning desks with visio sensors, which can be used to evaluate the health-related effects of sitting posture based on three dimensions, namely, human posture, critical threshold, and abnormal posture duration, are proposed. In accordance with the joint point model obtained from the OpenPose algorithm, we identified five abnormal sitting postures, namely, head tilt, body tilt, head lowering, reading at a close distance, and sitting for a long duration. Using this information, we designed a detection process for these postures to improve the accuracy of posture evaluation in our intelligent learning desks. We optimized the OpenPose model for mobile terminals by utilizing deep separable convolution to replace some convolution cores in the two-branch multistage network. This approach effectively reduced the amount of network structure parameters and significantly decreased the computational load required for the model. As a result of this optimization, we were able to more than double the video recognition speed compared with the original model. This improvement enables our intelligent learning desks to operate with greater efficiency on mobile devices without sacrificing accuracy or performance. According to our experiments and practical tests, our system can effectively monitor and warn students of common abnormal sitting postures. The recognition rate of abnormal sitting postures, such as prolonged learning, head tilt, body tilt, and head bow, has been optimized to over 92%. This high level of accuracy enables our intelligent learning desks to provide timely feedback and alerts to students when they exhibit poor posture habits, which can help prevent long-term health issues associated with prolonged sitting or incorrect posture.