Transfer Learning in Fall Detection Using Visual Data

Humans’ fall detection is critical in the prevention of life-threatening conditions. This is especially important for elderly people who are living alone. In this paper, we propose an approach based on a deep neural network for fall detection using data from the video camera and transfer learning techniques. We have chosen AlexNet as the baseline convolutional neural network architecture. The model was fine-tuned so it could solve a binary classification task (fall/non-fall). We used Le2i fall detection dataset, which was gathered in four locations with nine volunteers. 58 videos were used for model training and 47 videos were used for validation. The classifier has an accuracy and a Cohen’s kappa of 0.96 and 0.90, respectively, for the dataset where locations were familiar to the convolutional neural network. For the test dataset with previously unknown conditions, the accuracy and Cohen’s kappa were 0.96 and 0.80, respectively. Additionally, the heuristic application allowed us to increase specificity, that is, to decrease the rate of false alarms. This result may be potentially used as a part of a sensor-fusion fall detection system.