Infant Reaching Action Recognition in Unconstrained Environments
Action recognition can play a key role in the automation of devices, such as those developed for physical rehabilitation. The majority of current work focuses on adult action recognition, and only a limited body of action recognition work is geared toward pediatric populations, such as infants. This work introduces a new lightweight neural network structure, BabyNet, that recognizes reaching motion of infants from off-body stationary cameras. This approach makes use of the spatial and temporal connection between bounding boxes around an infant’s hands and object of interest in order to recognize a reaching action toward that object. BabyNet is trained and tested on a new dataset that showcases reaches performed in a sitting position by different infants in unconstrained environments, such as the families’ homes. To evaluate the efficacy of the proposed approach, an ablation study is conducted where BabyBet is compared against several other learning-based architectures. Results show that BabyNet performs satisfactorily in terms of average testing accuracy by exceeding that of competing networks. Due to its small size, it can serve as a lightweight architecture for video-based infant reaching action recognition.