Robust Pose-Based Human Fall Detection Using Recurrent Neural Network

Detecting falling event from the video for providing timely assistance to the fallen person is a challenging problem in computer vision due to the absence of large-scale fall dataset and the presence of many covariate factors like varying view angle, illumination, and clothing. In this paper, to address this problem, an effective approach for fall detection has been proposed. We have developed a recurrent neural network (RNN) with LSTM architecture that models the temporal dynamics of the 2D pose information of a fallen person. Human 2D pose information, which has proven effective in analyzing fall pattern as it ignores people's body appearance and environmental information while capturing the true motion information makes the proposed model simpler and faster. Experimental results have verified that our proposed method has achieved 99.0% sensitivity on both of the benchmark datasets of fall detection FDD and URFD.