Fall Detection among Elderly Person using FallCNN and Transfer Learning Models

Vergin Raja Sarobin M^*Jishnuraj KJANI ANBARASI LGraceline JasmineRukmani Panjanathan

• VIT University Chennai, Chennai, India

As per the data provided by the World Health Organization (WHO), falls are one of the major reasons for unintentional deaths or injuries in elderly people. Even though there are a lot of fall detection methods and algorithms exist, there is no efficient artificial intelligence strategy for detecting falls. Various literature states that Fall Detection among Elderly Person (FDEP) provides the possibility of bringing up an efficient and cost-effective way to tackle this problem. This paper generated a signal-based image dataset, SimgFall from the existing accelerometer or gyroscope-based sensor data of the SiSFall dataset for early detection of fall to fasten the medical assistance process. The SimgFall dataset is utilized to train and evaluate 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 for efficiently extracting diversified features from the SimgFall dataset. 1992 signal-based images of which 498 are the samples collected for fall, jump, stumble and walk of four classes respectively. The initial architecture referred as FallCNN_1, with two basic convolutional layers and max-pooling which is a simple and efficient in feature extraction and dimensionality reduction, resulted with 94% accuracy for detecting 4 classes. The incorporation of average pooling and dropout layers in FallCNN_2 reduced overfitting and improved feature extraction, enhancing accuracy to 95%. Expanding feature dimensions in FallCNN_3 further refined the model's capacity 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 the cumulative performance improvements through this architectural enhancement. Further, performance evaluation on the generated dataset using different pre-trained and custom models has been analysed based on the loss and accuracy curve. The experimental results show that the highest classification accuracy as 98% with a loss of 0.0833, utilizing categorical cross-entropy as the loss function.