Lightweight Video-Based Fall Detection

Abstract

Interactive presence monitoring has become increasingly popular for smart homes, with fall detection as one important application of interactive presence monitoring due to the medical benefits, which makes the topic of this research. Moreover, video data is employed as the same camera can be used for other wellbeing and security benefits. However, as the video-based monitoring raises privacy concerns, this research focuses on lightweight models to help them run locally on edge devices. Furthermore, because of the medically sensitive topic, the aim is to make the models transparent with explainable AI. EfficientNet is adopted as the backbone to learn spatial information from the video data by processing the video as a batch. LSTM is selected as the head to process the temporal data from the embeddings by reconstructing them to their original shape. This hybrid model is trained and evaluated on a narrowed-down version of the Omnifall benchmark with the cs-staged and cs-staged-wild splits. In addition, EfficientNet’s ability to learn spatial information is analysed with Score-CAM during inference, and the entire model is interpreted with t-SNE during testing. The benchmarks illustrate excellent results for the cs-staged splits, but inadequate results for the cs-staged-wild split, likely due to clean and laboratory environmental staged data. The evaluation is backed by t-SNE visualisations. Score-CAM illustrates the concerns about the dataset being too clean with significant biases during inference. The code for the research is available on GitHub: https://github.com/jesseValkama/Lightweight-Video-Based-Fall-Detection