Modern education has undergone tremendous progress, and a large number of advanced devices and technologies have been introduced into the teaching process. We explore the application of artificial intelligence to education, using AI devices for classroom behavior analysis. Embedded systems are special-purpose computer systems tailored to an application. Embedded system hardware for wearable devices is often characterized by low computing power and small storage, and it cannot run complex models. We apply lightweight models to embedded devices to achieve real-time emotion recognition. When teachers teach in the classroom, embedded portable devices can collect images in real-time and identify and count students’ emotions. Teachers can adjust teaching methods and obtain better teaching results through feedback on students’ learning status. Our optimized lightweight model PIDM runs on low-computing embedded devices with fast response times and reliable accuracy, which can be effectively used in the classroom. Compared with traditional post-class analysis, our method is real-time and gives teachers timely feedback during teaching. The experiments in the control group showed that after using smart devices, the classroom teaching effect increased by 9.44%. Intelligent embedded devices can help teachers keep abreast of students’ learning status and promote the improvement of classroom teaching quality.