Volume 13 | Issue 4
Volume 13 | Issue 4
Volume 13 | Issue 4
Volume 13 | Issue 4
Volume 13 | Issue 4
For the elderly, falls pose a serious public health danger on a global scale. If a fall is not prevented in time, it may significantly reduce an elderly person's mobility, independence, and quality of life. Battery life, user discomfort, expense, complexity of installation, furniture occlusion, and computational demands are only some of the issues with today's detection technologies. In this chapter, we offer a sensor-free method of detecting falls with the use of computer vision and applied machine learning techniques. We used OpenPose real-time multi-person 2D pose estimation to track an object's motion across two datasets totaling 570×30 frames captured in five rooms from eight vantage points[3,4]. A total of 13 joint points in the human leg are retrieved by the device, and any shifts in those points indicate human movement. The system is accurate in its identification of human joints and is able to filter out background noise to do so. By focusing on joint points rather than images, we may reduce training time and avoid the drawbacks of conventional image-based methods such as motion blur, lighting, and shadows. The usage of single-view images in this work helps to cut down on costly machinery. To study the dynamic changes in human joint points over time, we tried out time series recurrent neural network (RNN), long- and short-term memory (LSTM), and convolution neural network (CNN) models. The experimental findings demonstrate that the suggested model has superior fall detection accuracy to the state-of-the-art methods.