Low Complexity Fall Detection Using 2D Skeletal Data

Falling remains one of the events that can potentially cause irreversible damage, especially among elderly people if not treated early. Fall detection systems help to identify if a person has fallen and promptly alert the people involved to ensure timely medical care. Existing fall detection systems are found to utilize specialized sensors or require heavy computation requirements to identify the occurrence of falls. An easy-to-adopt, fall detection system is presented which can identify falls from a commodity-grade 2D CCTV camera. 2D skeletal data are extracted from the video frames and representative features are computed. The features specific to fall detection are grouped and analyzed in terms of their discriminative ability to distinguish falling from day-to-day actions. LSTM-based classifiers are trained on the filtered fall features to identify a person falling. The results are validated on the NTU (RGB) skeletal dataset. The precision and recall of the Fall detection classifier are found to be 98% and 100% respectively. The number of parameters in the developed system is found to be several magnitudes lesser than in existing systems, thereby validating low complexity with high accuracy.