The elderly population is increasing rapidly around the world. There are no enough caretakers for them. Use of AI-based in-home medical care systems is gaining momentum due to this. Human fall detection is one of the most important tasks of medical care system for the aged people. Human fall is a common problem among elderly people. Detection of a fall and providing medical help as early as possible is very important to reduce any further complexity. The chances of death and other medical complications can be reduced by detecting and providing medical help as early as possible after the fall. There are many state-of-the-art fall detection techniques available these days, but the majority of them need very high computing power. In this paper, we proposed a lightweight and fast human fall detection system using pose estimation. We used `Movenet' for human joins key-points extraction. Our proposed method can work in real-time on any low-computing device with any basic camera. All computation can be processed locally, so there is no problem of privacy of the subject. We used two datasets `GMDCSA' and `URFD' for the experiment. We got the sensitivity value of 0.9375 and 0.9167 for the dataset `GMDCSA' and `URFD' respectively. The source code and the dataset GMDCSA of our work are available online to access.