Multi-modal fall detection using improved YOLOv5 and mediapipe

Automated fall detection is critical for elderly care and health monitoring, yet detecting falls accurately remains challenging due to the variability in body orientations, occlusions, and lighting conditions. To address these challenges, this paper presents an enhanced fall detection system based on YOLOv5s and MediaPipe, tailored for improved feature extraction and spatial refinement. Key contributions of our approach include the integration of the Dynamic Atrous Convolution (DAC) and Self-Calibrated Convolution (SCConv) modules into YOLOv5's backbone and head, enhancing multi-scale feature extraction and spatial awareness. The system leverages both MediaPipe's precise keypoint detection and YOLOv5's robust object localization to calculate metrics such as descent velocity and bounding box aspect ratio for reliable fall detection. We evaluate our model using public datasets, demonstrating its robustness across varied scenarios. Experimental results show that our enhanced YOLOv5 model achieves a significant improvement in mean Average Precision (mAP), reaching up to 98.6%, a 4.8% increase from original yolov5s and outperforming conventional methods. The proposed approach offers a promising solution for enhancing the accuracy and reliability of automated fall detection systems, which is crucial for improving elderly healthcare and monitoring.