This paper presents a computer vision-based method for assessing and predicting autism spectrum disorders (ASDs).Traditional diagnostic methods for ASD have disadvantages such as high cost, strong subjectivity, and long time, but thismethod is convenient, low cost, and avoids some defects of traditional diagnostic methods. The research team designed asocial interaction scenario and collected an autistic child Face dataset (ACFD), then used computer vision methods to extractinformation about the children’s faces. On this basis, multiple features of children were obtained from four aspects: facialappearance time, eye concentration analysis, response time to name calls, and emotional expression ability, which were usedto assess and compare children with autism and typical development (TD). Finally, multiple features were linked together andmachine learning methods were used to classify children. The experimental results show that multiple features can reflect thetypical symptoms of children, and the classification accuracy of the prediction model is up to 92.16%, which proves that theautomatic recognition method can provide auxiliary diagnosis and data support for doctors. This method provides a new ideaand direction for early diagnosis and intervention of ASD, which is of great significance for improving the quality of life andtreatment effect of children with ASD.