Patients admitted to intermediate care units require frequent monitoring by hospital personnel. An automatisation of this monitoring would save a considerable amount of resources and could also improve the quality of the treatment. In this thesis, a deep learning-based video action recognition model is through different transfer learning approaches trained to distinguish between behaviours of patients in TV-series and a prediction system which collects, processes and predicts on images in real-time is proposed. The results from the model-training suggest that it is possible to detect behaviours that need human intervention but training on a large-scale, real-life dataset is required to form a solid conclusion. The performance results of the prediction system show that live-streamed predictions are possible at frame rates sufficient for capturing sought events, without GPU acceleration. / Patienter inlagda på intermediärvårdsavdelningar behöver frekvent övervakning av sjukhuspersonal. En automatisering av denna övervakning skulle spara en betydande mängd resurser och även kunna förbättra kvaliteten av behandlingen. I detta examensarbete tränas en djupinlärningsbaserad modell för videohandlingsigenkänning att, genom olika överföringsinlärningsmetoder, skilja på beteenden mellan olika patienter i TV-serier och ett prediktionssystem som insamlar, processerar och predikterar på bilder i realtid presenteras. Resultaten från modellträningen tyder på att det är möjligt att detektera beteenden som kräver mänsklig interaktion men träning på ett storskaligt, realistiskt dataset krävs för att kunna dra en säker slutsats. Prestandaresultaten från prediktionssystemet visar att live-strömmade prediktioner är möjliga vid bilduppdateringsfrekvenser tillräckliga för att fånga de sökta händelserna, utan GPU-acceleration.

Static image action recognition aims at identifying the action performed in a given image. Most existing static image action recognition approaches use high-level cues present in the image such as objects, object human interaction, or human pose to better capture the action performed. Unlike images, videos have temporal information that greatly improves action recognition by resolving potential ambiguity. We propose to leverage a large amount of readily available unlabeled videos to transfer the temporal information from video domain to static image domain and hence improve static image action recognition. Specifically, We propose a video prediction model to predict the future video of a static image and use the future predicted video to improve static image action recognition. Our experimental results on four datasets validate that the idea of transferring the temporal information from videos to static images is promising, and can enhance static image action recognition performance. / Master of Science / Static image action recognition is the problem of identifying the action performed in a given image. Most existing approaches use the high-level cues present in the image like objects, object human interaction, or human pose to better capture the action performed. Unlike images, videos have temporal information that greatly improves action recognition. Looking at a static image of a man who is about to sit on a chair might be misunderstood as an image of a man who is standing from the chair. Because of the temporal information in videos, such ambiguity is not present. To transfer the temporal information and action features from video domain to static image domain and hence improve static image action recognition, we propose a model that learns a mapping from a static image to its future video by looking at a large number of existing images and their future videos. We then use this model to predict the future video of a static image to improve its action recognition. Our experimental results on four datasets show that the idea of transferring the temporal information from videos to static images is promising, and can enhance static image action recognition performance.