To address the limitations of traditional fall detection methods, which are prone to false detections under occlusion and illumination interference and struggle to balance lightweight design with detection accuracy, this paper proposes an improved YOLOv8-based algorithm for indoor elderly fall detection. Specifically, omni-dimensional dynamic convolution is integrated into the C2f module of the backbone network, enabling adaptive feature extraction and enhancing representational capability. In the neck network, the C2f module is further optimized with the FasterNet module to effectively reduce computational cost. In addition, a large selective kernel attention (LSKA) mechanism is embedded into the SPPF module to improve detection precision, while a spatial-enhanced attention module (SEAM) is introduced into the detection head to further strengthen discriminative ability. Comparative and ablation experiments were conducted, and the detection results were further visualized as heatmaps to validate the effectiveness of the proposed approach. Experimental results demonstrate that, compared with the baseline YOLOv8n model, the improved algorithm achieves an mAP@0.5 of 91.0% (an improvement of 1.1%), with a 20.6% reduction in parameters and a 42.7% decrease in GFLOPs, thereby confirming that the proposed method effectively balances detection accuracy and lightweight design in indoor elderly fall detection tasks.