关键词: 注意缺陷多动障碍, 动作识别, 骨架数据, 图卷积网络；3D卷积神经网络

Abstract: Attention Deficit Hyperactivity Disorder (ADHD) is a common neurodevelopmental disorder in childhood, characterized by inattention, hyperactivity, and impulsivity, often exhibiting specific motion patterns. Traditional action recognition algorithms were found to have low recognition accuracy and slow response speed when handling these specific actions. To address such issues, a method was adopted where 2D pose data were transformed into 3D heatmaps, and spatial relationships between joints were precisely represented using Gaussian distribution, which effectively preserved spatiotemporal information. To overcome the limitations of single-modal data, a multimodal fusion method based on skeleton and 3D heatmaps was introduced. The output features of Short 3D-CNN and Adaptive Graph Convolutional Network (AGCN) were fused to fully exploit the advantages of both modalities, thereby improving action recognition performance. Experimental results on the ADHD patient dataset collected at the Mental Health Center of West China Hospital, Sichuan University, showed that the proposed algorithm achieved a TOP-1 recognition accuracy of 0.8604 and a TOP-5 recognition accuracy of 0.9873 for eight different types of actions. Additionally, an automatic ADHD classification algorithm based on action types was proposed, which classified ADHD into head and facial movement type, trunk movement type, and limb movement type, achieving a recognition accuracy of 0.75 and a response time of 5 seconds. Compared with 2s-AGCN and POSEC3D, the proposed algorithm demonstrated higher recognition accuracy in complex action scenarios, providing a new technical approach for personalized ADHD intervention.

Key words: Attention Deficit Hyperactivity Disorder (ADHD), action recognition, skeleton data, Graph Convolutional Network (GCN), 3D Convolutional Neural Network (3D-CNN)