Abstract

The domain of automatic emotion recognition introduces innovative methods and technologies that hold potential for enhancing therapy for children with autism. This research paper aims to explore various approaches and tools utilized to recognize emotions in children. The study presents a comprehensive literature review conducted using a systematic approach and the PRISMA methodology for reporting both quantitative and qualitative findings. The analyzed studies employ diverse observation channels and modalities, such as facial expressions, prosody of speech, and physiological signals. Among the recognized emotions, the basic ones, including happiness, fear, and sadness, are the most frequently identified. Both single-channel and multichannel approaches are used, with a preference for the former. For multimodal recognition, early fusion emerges as the most commonly applied technique. Building classifiers, Support Vector Machines (SVM), and neural networks are the prevailing methods. Through qualitative analysis, significant insights are gained concerning participant group construction and the most prevalent combinations of modalities and methods. However, all channels are reported to be susceptible to some disturbances, leading to temporary or permanent unavailability of specific emotional symptom information. Furthermore, the paper identifies challenges in devising appropriate stimuli, labeling methods, and creating open datasets for research in this field. Addressing these challenges could foster further advancements in the therapeutic support for children with autism.