An efficient algorithm for pedestrian fall detection in various image degradation scenarios based on YOLOv8n.

With the growing elderly population, especially in urban areas, pedestrian fall detection for the elderly has become a critical global concern. Existing pedestrian fall detection systems often suffer from low accuracy and poor performance under challenging conditions such as rain, snow, nighttime, or camera obstructions. To address these issues, this paper proposes an enhanced pedestrian fall detection algorithm called Pedestrian Chain-of-Thought Prompted Adaptive Enhancer YOLO (PCE-YOLO), based on YOLOv8n. Several improvements were made to YOLOv8n, including the integration of a Chain-of-Thought Prompted Adaptive Enhancer (CPA-Enhancer) module to boost detection performance in complex environments. Additionally, the Cross Stage Partial Bottleneck with 2 Convolution Block (C2f) was optimized to reduce computational load and parameter count without compromising performance, while the Inner Extended Intersection over Union (Inner-EIoU) loss function was employed to improve bounding box regression accuracy and speed. To validate the model's effectiveness, a dataset of 7,782 pedestrian fall images was collected, and three degraded image datasets were generated to simulate real-world conditions. PCE-YOLO improved the mean Average Precision (mAP) by 4.52% compared to YOLOv8n on both original and degraded datasets, respectively. Moreover, it achieved a frame per second (FPS) rate of 210.5, making it suitable for real-time detection applications. The results demonstrate that PCE-YOLO significantly enhances detection accuracy and speed in various challenging environments, offering a robust solution for real-time pedestrian fall detection.