Efficient Service Provisioning in Fog Computing for Multimodal Fall Detection Systems
Sarangapani Nivarthi, Vijaya N. Aher, T. Kuppuraj, Ankit Punia, Rajiv Gandhi N, Lalit Khanna · 2024
Our capstone project makes use of an actual-world app that can be efficiently managed by an Edge-Fog-Cloud system. Fall detection is an issue that we have taken up and aim to solve by providing decisions with the lowest feasible delay. We propose a three-stage process that makes use of multimodal data (sensors, video, and pictures). An edge device's convolutional neural network infers the visual data used for the first inference. To improve weight distribution and training speed, the neural network is constructed utilizing transfer learning. The optical flow method removes the background noise and focuses just on the individual by representing moving objects as vector fields. A condensed neural network is able to operate on a resource-constrained edge device. After that, the sensor data is classified by the fog layer using SVM, Random Forest (RF) and Gradient Boosting (XGBoost). The fog layer also uses visual and sensor data to make the ultimate fall judgment. The ensemble decision-making process relies on the inferences drawn from the fog and edge layers to arrive at this conclusion. Model construction and durable data storage are carried out at the cloud layer. For more accurate evaluations, our work also deals with false negative rates. To do this, we use augmentation of images on the misclassified photos; this improves the classification and adds value to the dataset. Last but not least, the suggested models are containerized for simple deployment, maintenance, and scalability. With an accuracy of 99%, the proposed technique surpassed all of its competitors, including SVM, XGBoost. Every approach has its advantages, but the Proposed technique clearly outperforms the others and is hence the best option for this specific job.