Existing pre-impact fall detection systems have high accuracy, however they are either intrusive to the subject or require heavy computational resources for fall detection, resulting in prohibitive deployment costs. These factors limit the global adoption of existing fall detection systems. In this work we present a Pre-impact fall detection system that is both non-intrusive and computationally efficient at deployment. Our system utilizes video data of the locality available through cameras, thereby requiring no specialized equipment to be worn by the subject. Further, the fall detection system utilizes minimal fall specific features and simplistic neural network models, designed to reduce the computational cost of the system. A minimal set of fall specific features are derived from the skeletal data, post observing the relative position of human skeleton during fall. These features are shown to have different distributions for Fall and non-fall scenarios proving their discriminative capability. A Long Short Term Memory (LSTM) based network is selected and the network architecture and training parameters are designed after evaluation of performance on standard datasets. In the Pre-impact fall detection system the computation requirement is about 18 times lesser than existing modules with a comparable accuracy of 88%. Given the low computation requirements and higher accuracy levels, the proposed system is suitable for wider adoption in engineering systems related to industrial and residential safety.