Leonardo Camelo

The early identification of anomalous movements in infants is crucial for intervening in potential neuromotor development disorders. The clinical method General Movement Assessment (GMA) is devoted to this identification task. However, since GMA is intensive and requires experts, new machine learning-based approaches and keypoints extracted from videos have emerged. However, challenges such as the underrepresentation of infants with writhing movements (WM)—general movements presented by infants in their first weeks of life; the scarcity of public datasets; and the fact that only video segments showing infants performing movements must be analyzed, are limitations to identify anomalous movements in infants automatically. This work introduces a method which uses spatial distance features extracted from skeletal data and employs subspace method based on the statistical analysis of the eigenvalue-entropy to enhance the detection of infants movements in video data, especially video from infants exhibiting WMs. The proposed method applies a subspace approach as an initial step to filter infant movements for further detection and subsequent classification, aiming to improve the detection and understanding of these critical early indicators. The results show that the proposed method is able to detect subtle nuances in infant movements more effectively than the baseline method, making it a promising tool for automatic developmental monitoring.