Bi-Modal Multiperspective Percussive (BiMP) Dataset for Visual and Audio Human Fall Detection

Falls are the leading cause of injuries and fatalities among older individuals propagating concerns for health and safety. Mitigating these concerns requires timely intervention and response to minimize health complications. Autonomous fall detection systems are often offered as a means to alleviate these concerns. Typical fall detection systems classify a fall event using either inertial- or vision-based data. Aside from these two modes of input for fall detection, sparse human fall audio data is available. Audio datasets associated with the dynamic impact of human falls from the perspective of a spatial environment are both unavailable publicly and focused on niche applications. This dataset provides multimodal composition of visual and audio data to aid in fall detection technique development. Though this dataset emphasizes audio novelty, a synchronized visual component is included offering a multifaceted collection of data. Audio input for fall detection systems presents an opportunity for new approaches towards autonomous fall detection. This dataset comprises human movement representing activities of daily living and falls of 25 participants in various residential settings, yielding 1,300 instances of unique visual and audio samples. Promising utility of the bi-modal multiperspective percussive (BiMP) dataset is demonstrated through experimental data evaluations using techniques including: GoogLeNet, Long Short Term Memory, Continuous Wavelet Transforms, and Short-time Fourier Transforms for human fall detection achieving accuracies up to 96%. The concept of audio-based fall detection has the potential to mitigate concerns regarding privacy and invasiveness, in addition to broadening the scope of fall detection mechanisms.