Abstract
Autism Spectrum Disorder (ASD) is a neurodevelopmental disorder characterized by differences in social communication and repetitive behavioral patterns. Behavioral analysis based on video is commonly performed through manual observation, which can be time-consuming and subjective. This study proposes a CNN-LSTM Autoencoder approach to detect behavioral anomalies in children with ASD based on video data. The CNN Autoencoder extracts spatial features from video frames, while the LSTM Autoencoder learns temporal patterns from sequential features. The model was trained using 88 videos collected from children with ASD, while five videos from a public dataset were used for testing. The experimental results produced mean reconstruction errors of 0.1000, 0.1184, and 0.2079 for the training, validation, and testing data, respectively, with a threshold of 0.2520. Among 65 testing sequences, 51 were classified as in-distribution with respect to the ASD training data, while 14 were identified as anomalies. These results indicate that the proposed approach can identify behavioral deviations from the learned ASD behavioral distribution. However, the system is not intended to provide a clinical diagnosis of ASD.