Abstract: Autism Spectrum Disorder (ASD) is characterized by repetitive behaviors such as spinning, head banging, and arm flapping. Early detection and classification of these behaviors can aid in therapeutic interventions. This paper presents a deep learning-based approach for recognizing and classifying autistic behaviors in video data. We trained a model using a dataset of 12 training videos and 3 validation videos, achieving an F1-score of 0.8342. This model uses a sequence transformer-based neural network to achieve highly accurate behavior classification. Our results demonstrate strong performance, particularly in spinning detection (AUC = 0.99), while arm flapping recognition shows room for improvement (AUC = 0.83). The system provides real-time analysis with an average inference time of 0.58 seconds per video, making it suitable for clinical and assistive applications.

Keywords: Autism Spectrum Disorder (ASD), deep learning, behavior recognition, computer vision, repetitive behaviors

Pages: 206-215