Deep Learning Based Real Time Fall Detection Model H Aravind Sarma Department of Electronics and Communication Mar Baselios College of Engineering and Technology Trivandrum, India aravindsarmah@gmail.com Jijo Jose Department of Electronics and Communication Mar Baselios College of Engineering and Technology Trivandrum, India jijo.jose@mbcet.ac.in Abstract—Health professionals are enormously utilising the advantages of most advanced technologies, leading to a scalable development in the field of health care. There has been a paradigm shift from manual monitoring towards more accurate virtual monitoring with a high accuracy rate. The world is very much concerned about taking care of old people. One of the worst things that can happen to older individuals is a fall. The creation of fall detection systems is urgently needed due to the aging population's rapid growth. With the development of technology, fall detection and rescue systems are able to provide quick emergency assistance to victims of accidental falls, thereby reducing the number of fatalities by detection of fall and providing adequate care. Here, we propose a model for real-time fall detection that uses deep convolutional neural networks. An automated fall detection model is essential for elderly care, since it offers a real time monitoring and fall detection whereas manual fall detection and immediate attention depends on timely availability of many factors. A real time monitoring system coupled with Deep learning techniques can accomplish this task. For fall detection, many deep learning networks can be utilized. Here, we propose an efficient model for real-time fall detection that uses deep convolutional neural networks and torch tensors. The system will recognize Fall and Non Fall from a set of real time input videos and respond according to scenario. We tried a hardware demonstration of an Automated Fall Detection System in an Up Extreme Series single board computer with LSTM neural network model. Keywords—Deep Learning, Convolutional Neural Networks, Pytorch. I. INTRODUCTION A fall is a wild drop of the body to the ground level. This criterion has been the primary source of inspiration for the majority of studies on the development of fall detection devices. The majority of studies agreed that aging, brittle bones, diseases, alcohol use, poor vision, slick flooring, inadequate lighting, and a loss of body balance are common risk factors. One of the central point that add to falls is dizziness. Four phases make up the time or event of the fall ● Pre-fall phase: Behaviors such as abruptly sitting down, running and walking are conducted during this period. ● Critical phase: Within a couple of seconds, the body begins to move erratically and falls to the ground. ● Post-fall phase: Body remains mostly still and on the floor or ground. Reduce this period as much as possible to lessen the damage that happened during the fall. ● Recovery phase: The body returns to the prefall phase either naturally or with the assistance of caretakers. The core principle of the system is the identification of the critical window or post-fall phase. The accuracy with which falls are detected is significantly impacted by the pre-fall period. Activities that are routine, like abruptly sitting down in chairs or deliberately laying down on the ground, are additionally erroneously distinguished as falls. A. Universal Fall Detection System Figure 1: Universal Fall Detection System Four key tasks are carried out by a fundamental fall detection system: sensing or data acquisition, data processing, classification, and communication. The first and most crucial phase is called sensing or data acquisition. This involves using the appropriate sensors, such as an accelerometer, gyroscope, pressure sensor, sound sensor, camera, radar, or other sensors, to detect the appropriate physical quantity. Data processing, such as filtering, amplification, background reduction, and so on, is the second step. The third and most crucial step is the classification of events and the extraction of significant features. The classifier algorithm determines whether a fall has occurred by separating the significant information from the processed data. For computational activities like feature extraction and classification various microcontrollers and laptops or personal computers are used over a wireless network. Wireless communication protocols are employed to analyse the data from sensors remotely. Based on the results of the classifier, the system's final step generates an alarm to notify the person who should know. It can be extended to creation of a pool of fall detection systems with medical emergency services so that this automated system can ensure a timely medical response on detection of fall.