YOLOv8-based recognition of preschool children's interactive behaviors during picture book reading with an improved attention mechanism

With the continuous evolution of intelligent education technology, the automatic recognition of preschool children's picture book reading interaction behavior has become an important research direction for improving the objectivity of educational assessment and the timeliness of intervention. Traditional manual observation methods are difficult to achieve continuous, fine-grained analysis of behavior in natural scenes. In contrast, methods based on YOLOv8 (You Only Look Once version 8) and improved attention mechanisms provide an effective solution for visual-based interactive behavior recognition. This study constructs a labeled dataset containing ten types of behaviors, including gaze, pointing, and book turning, and integrates the CBAM (Convolutional Block Attention Module) attention mechanism into the YOLOv8 network to enhance the perception of key interactive features. Experimental results show that the overall recognition accuracy of the improved model reaches 91.3%, an improvement of 13.1% compared with the basic YOLOv8; the accuracy of gaze at picture books, pointing, and verbal questioning behaviors are 93.5%, 89.1%, and 87.4%, respectively, and the recall rates are improved to 91.2%, 86.4%, and 84.1%, respectively. At the same time, the overall false-positive rate of the model is reduced to 6.3%, the processing time per frame is 32.1 milliseconds, and the real-time satisfaction rate is 85.7%, which achieves a balance between processing efficiency and recognition accuracy. This research provides an experimental basis for automated behavior analysis. The proposed method differs from the original CBAM by introducing multi-scale adaptive channel attention and local-global spatial fusion.