Fall detection among elderly persons using FallCNN and transfer learning models

According to data provided by the World Health Organization (WHO), falls are one of the major reasons for unintentional deaths or injuries in older adults. Although many fall detection methods and algorithms exist, there is no efficient artificial intelligence strategy for fall detection. Various studies have stated that Fall Detection among Elderly Persons (FDEP) provides the possibility of developing an efficient and cost-effective way to tackle this problem. This study generated a signal-based image dataset, SimgFall, from the existing accelerometer or gyroscope-based sensor data of the SiSFall dataset for the early detection of falls to accelerate the medical assistance process. The SimgFall dataset was used to train and evaluate the FallCNN model, a novel deep Convolutional Neural Network (CNN) architecture comprising multiple CNN folds to effectively learn discriminative features from the transformed signal representations. These models utilize depth-wise convolution with varying dilation rates to efficiently extract diversified features from the SimgFall dataset. The dataset contained 1992 signal-based images, of which 498 were samples collected for fall, jump, stumble, and walk for the 4 classes. The initial architecture, referred to as FallCNN_1, with two basic convolutional layers and max-pooling, which is simple and efficient in feature extraction and dimensionality reduction, resulted in 94% accuracy for detecting the 4 classes. The incorporation of the average pooling and dropout layers in FallCNN_2 reduced overfitting and improved feature extraction, thereby enhancing the accuracy to 95%. Expanding the feature dimensions in FallCNN_3 further refined the capacity of the model to capture intricate patterns, achieving a notable accuracy of 97%. Finally, FallCNN_4 with three convolutional blocks and additional intermediate layers achieved the highest accuracy of 98%, demonstrating cumulative performance improvements through architectural enhancements. Furthermore, the performance of the generated dataset using different pretrained and custom models was evaluated based on the loss and accuracy curves. The experimental results showed that the highest classification accuracy was 98%, with a loss of 0.0833, using categorical cross-entropy as the loss function.