Detecting and alerting for falls is a crucial component of both healthcare and assistive technologies. Wearable devices are vulnerable to damage and require regular inspection and maintenance. Manned video surveillance avoids these problems but it involves constant labor-intensive attention and, in most cases, may compromise the privacy of the observed individuals. To address the issue, in the work we introduce and evaluate a novel approach to fully automated fall detection. The presented technique uses direct reconstruction of principal motion parameters, avoiding the computationally expensive full optical flow reconstruction and still providing relevant descriptors for accurate detections. Our method is systematically compared with state-of-the-art techniques. Comparisons of detection accuracy, computational efficiency, and suitability for real-time applications are presented. Experimental results demonstrate notable improvements in accuracy while maintaining a lower computational cost compared to traditional methods, making our approach highly adaptable for real-world deployment. The findings highlight the robustness and universality of our model, suggesting its potential for integration into broader surveillance technologies. Future directions for development will include optimization for resource-constrained environments and deep learning enhancements to refine detection precision.
Karpuzov et al. (Fri,) studied this question.