Vision-based techniques for fall detection in 360∘ videos using deep learning: Dataset and baseline results

Alarming cases of falls in the elderly have triggered the rise of robust and cost-efficient systems for automated fall detection in humans. Although several potential solutions exist, they still have not achieved the desired level of robustness and acceptability. Lately, the proliferation of low-cost cameras coupled with deep learning techniques has transformed vision-based methods for fall detection. Motivated by this, in this paper, we present an alternate low-cost and efficient system for fall detection in 360 ∘ videos using deep learning. Towards this, we first built a well-balanced video dataset named Fall360. The Fall360 dataset contains video clips of several falls and non-fall actions, captured by a 360 ∘ camera mounted on the ceiling in a home-like environment. Secondly, we examined the performance of deep learning techniques that consist of several variants of hybrid CNN & LSTM, hybrid CNN & ConvLSTM, and 3D CNNs to test the effectiveness of the dataset in the fall detection task. Thirdly, to assess the performance of these techniques, we conducted an ablation study on a recently introduced multi-camera UP-Fall dataset. The deep learning models attained substantial improvement in recognition accuracy on both the fall datasets and have set the new state-of-the-art performance. Overall, our designed fall detection system using 360 ∘ videos, in addition to providing a better perspective, bestows a more suitable and low-cost alternative for the existing multi-camera-based fall detection systems. To encourage more study, we will make our in-house Fall360 dataset publicly available to the research community.