Abstract:
Falls represent a substantial risk to the elderly and individuals with particular
health conditions, often leading to serious injuries, loss of independence, and even
mortality. Early detection of falls is critical for timely medical intervention and also
improves patient outcomes. However, traditional fall detection systems that rely on
wearable sensors can be intrusive, inconvenient, or unsuitable for continuous
monitoring. This system proposed a vision-based human fall detection alert system that
combines the YOLO (You Only Look Once) object detection network with a pose
estimation to recognize fall events effectively. In the preprocessing stage, input videos
are processed through frame extraction, resizing, normalization, and format conversion
to prepare the data for analysis.
The YOLOv8 pose estimation algorithm examines skeletal keypoints in video
frames with the height-width ratio and angle-based method to identify whether a human
falls or does not fall from other common human activities, such as walking, sitting, or
falling. Upon detecting a fall, the system immediately activated a voice-based alarm
notification using a pre-recorded AVI audio file, enabling rapid response and
assistance. The model has been trained and evaluated by using both the Le2i Fall
Dataset, M-Dataset, and CCTV-Dataset, ensuring robustness across different
environments. The experimental results showed that both angle-based and height-width
ratio methods performed well, with the overall accuracies ranging from 95% to 98%
and F1-scores between 94% to 97%. The proposed system provided a cost-effective
and accurate solution for real-time fall detection, demonstrating high potential for
integration into smart healthcare monitoring systems.