Person Fall Detection Using Weakly Supervised Methods

A fall can result in severe injuries and even fatalities. An automatic fall detection system could potentially save lives by alerting other people of the accident. Current approaches to fall detection systems include accelerometers and other physical sensors that have several drawbacks. Current computer vision-based approaches to fall detection are trained and tested on very simple and unrealistic datasets. Creating a new dataset for traditional supervised learning would require a significant amount of time for annotating the dataset. We, therefore, explore weakly supervised methods from the Video Anomaly Detection (VAD) literature and collect a new dataset to test the viability of a reliable fall detection algorithm using the VAD framework. We explore Multiple Instance Learning and propose a model with a novel loss function that outperforms state-of-the-art weakly supervised anomaly detection models in fall detection. Furthermore, our approach achieves competitive performance compared to the current state of the art in UCF-Crime despite being much simpler.