Multi Visual Modality Fall Detection Dataset

Falls are one of the leading causes injury-related deaths among elderly worldwide. Effective detection falls can reduce risk complications and injuries. Fall be performed using wearable devices or ambient sensors; these methods may struggle with user compliance issues false alarms. Video cameras provide a passive alternative; however, regular red, green, blue (RGB) impacted by changing lighting conditions privacy concerns. From machine learning perspective, developing an effective fall system is challenging because rarity variability falls. Many existing datasets lack important real-world considerations, such as varied lighting, continuous activities daily living (ADLs), camera placement. The considerations makes it difficult to develop predictive models that operate effectively in real world. To address limitations, we introduce novel multi-modality dataset (MUVIM) contains four visual modalities: infra-red, depth, RGB thermal cameras. These modalities offer benefits obfuscated facial features improved performance low-light conditions. We formulated anomaly problem, which customized spatio-temporal convolutional autoencoder was trained only on ADLs so would increase reconstruction error. Our results showed infra-red provided highest level (AUC ROC=0.94), followed ROC=0.87), depth ROC=0.86) ROC=0.83). This research provides unique opportunity analyze utility detecting home setting while balancing performance, passiveness, privacy.