Towards Automated Emotion Classification of Atypically and Typically Developing Infants

Abstract

The World Health Organization estimates that 15 million infants are born preterm every year [1]. This is of concern because these infants have a significant chance of having neuromotor or cognitive developmental delays due to cerebral palsy or other developmental issues [2]. Our long-term goal is to determine the roles emotion and movement play in the diagnosis of atypical infants. In this paper, we examine how automated emotion assessment may have potential to classify typically and atypically developing infants. We compare a custom supervised machine learning algorithm that utilizes individual and grouped facial features for infant emotion classification with a state-of-the-art neural network. Our results show that only three concavity features are needed for the concavity algorithm, and the custom algorithm performed with relatively similar performance to the neural network. Automatic sentiment labels used in tandem with infant movement kinematics would be further investigated to determine if emotion and movement are interdependent and predictive of an infant's neurodevelopmental delay in disorders such as cerebral palsy.

Figures

Fig. 1.

PANDA Gym

Fig. 2.

PANDA Camera Views

Fig. 3.

The BAEBIS Scale

Fig. 4.

Example FER 2013 and CIF Images

Fig. 5.

Custom-Made Algorithm Flowchart

Fig. 6.

68 keypoint extraction performed on sample CIF image (B1PosBW)

Fig. 7.

Function fit for top left eye of CIF image (B1PosBW). The concavity value is extracted from this functon.

Fig. 8.

Infant 18: Concavity Labels and Open Pose Velocity Tracking