Addressing the challenge of low accuracy in detecting classroom behaviors from teaching videos, compounded by the absence of a public classroom behavior dataset, this paper presents a novel approach. We construct a dataset specific to student classroom behavior and propose a detection method based on an improved SlowFast architecture. Our method integrates several innovations to enhance performance. Firstly, we introduce a pyramid segmentation attention module to replace the 3x3 convolution in the residual network. This module establishes a dependency relationship for long-term channel attention while capturing multi-scale spatial information. Secondly, we augment the fast branch of the network with a Transformer coding module to capture more timing information, thereby improving model accuracy. Furthermore, we modify the loss function of the original model to a dynamically scaled cross-entropy loss function. This adjustment reduces the loss weight of easily trainable samples, enabling the network to prioritize challenging samples and address the issue of unbalanced positive and negative samples. Experimental results on our self-constructed classroom behavior dataset demonstrate that the improved model achieves an average precision of 89.04%, a 3.58% improvement over the original model, and effectively reduces false detections.