Multi-Person Fall Detection Using Data Assimilation Method With Kalman Filter

Fall detection is an essential technology for ensuring the safety of elderly individuals, as falling accidents are critical and can cause significant functional damage in old age. Our previous work focused on fall detection with just a single individual, using simple statistical aggregation to achieve low model complexity and moderate accuracy. However, the pose estimation model may score low confidences on individual body parts (landmarks), affecting the aggregated statistics, thereby resulting in incorrect fall detection status. To solve this problem, we propose an enhanced fall detection method that adopts the Kalman filter for improved landmark and fall detection and object tracking for multi-person fall detection. Specifically with the Kalman filter, we reduce noise in network model’s heatmaps and landmarks with an adaptability across different input video sources. Compared to other methods that analyze AI models’ hidden layers and the layer outputs for providing confidence of measurements, this approach has a big advantage of plug-and-play for the pose estimation models and other streaming models that provide confidence. Tested on a computer with i7-13700HK, RTX 4070, 32 GB RAM, and a full HD camera, our method achieved an F1-score of 0.944 in a multi-person setup and 0.933 in the single-person setup.