Integrating CLAHE and H-MHI with CNN for prompt recognition in children with a diagnosis of autism

Abstract

Individualized therapeutic prompts such as physical, gestural, or verbal cues are necessary to enable children diagnosed with autism to become independent in daily activities. However, the automated detection of the prompts in naturalistic conditions is difficult. The solutions currently proposed also have difficulty with unstructured video data obtained from real-world therapeutic settings with different camera angles, uneven lighting, and random movement of children. Therefore, this study aims to introduce a new model that combines contrast limited adaptive histogram equalization (CLAHE) along with hybrid motion history image (H-MHI) to extract motion-based features with convolutional neural networks (CNNs) to classify features and provide timely responses in the context of autism therapy. The H-MHI method was based on the regular MHI with the addition of Otsu thresholding and Canny edge detector. These features were subsequently categorized with the CNN architectures in the form of VGG19 and MobileNetV2. The suggested algorithm was tested on 1,083 video samples and the results showed the ability of H-MHI to enhance accuracy by 2-3% at less computing time than the conventional MHI. The trend reflected the suggested method’s effectiveness in enhancing recognition performance without compromising the computational efficiency.