Due to uncertainties in indoor and outdoor settings, occlusions, and variations in human postures at specific angles, issues such as low recognition accuracy and poor real-time performance arise. This paper proposes an improved pedestrian fall detection algorithm based on YOLOv8. Firstly, the enhanced YOLO v8 algorithm is applied to detect falls, utilizing the SPD-Conv, a module composed of a space-to-depth (SPD) layer followed by a non-strided convolution (Conv) layer, to replace the traditional stride convolutions and pooling layers in the Backbone, enhancing the handling of small objects and low-resolution images. Then, the SlimNeck lightweight backbone network is employed as the structural base for the Neck, reducing the model’s parameter count and computational demand while maintaining detection performance. Finally, the original loss function is replaced with Inner-EIoU, introducing new geometric constraints to accelerate convergence and improve detection performance. According to comparative experimental results, this algorithm outperforms other current mainstream one-stage algorithms.