Fall-LSTM: A Fall Detection Network Based on Spatio-Temporal Location Module

Fall events have unique dynamic features, which are not fully utilized by existing fall detection methods. Based on video understanding, we propose Fall-LSTM to learn such features pertinently without additional inputs. Fall-LSTM is composed of CNN-LSTM framework and two excitation modules i.e., Spatial Attention Module (SAM) and Temporal Location Module (TLM). SAM provides spatial constraints on motion for feature layers through foreground extraction and spatial pooling. TLM emphasizes frames with high probability of fall events to LSTM by inferring the rate and trend of motion in clips. Experimental results show that our proposed modules significantly improve the performance of LSTM model, outperforming the state-of-the-art methods on two public Fall Detection Datasets i.e., Le2i and UR.