Fall-Detection Model for Elderly People to Improve Safety Based on Enhancing Inception v3 and LSTM

Abstract

Fall detection, especially among the elderly, is crucial due to the potentially severe consequences of the delayed identification of falls. This paper introduces an innovative approach to enhance the Inception v3 model. It involves adding the feature extractor ends with a global average pooling layer to address overfitting issues and eliminate the SoftMax layer. Additionally, LSTM is incorporated to improve classification accuracy by preprocessing the data and removing the background, thereby reducing confusion and boosting accuracy. The initial steps include obtaining a dataset from video footage, converting it to a series of images, and subjecting it to preprocessing for the next stages of training and testing. Our model was trained and tested using three distinct datasets: our dataset, the fall-detection dataset, and the Le2i dataset. The extracted features were fed into the Long Short-Term Memory (LSTM) classifier for fall detection classification. The LSTM classifier leverages these features to distinguish between fall and non-fall instances. The proposed model achieves significant accuracy, with scores of 0.98 on our dataset, 0.97 on the Le2i dataset, and 0.96 on the Fall Detection Dataset, underscoring its effectiveness in robust fall detection. This study contributes to fall detection, particularly in scenarios crucial to the well-being of the elderly population.