A mobile-based CMD-YOLO detection method was proposed in order to address the challenges of complex background, difficulty in detecting human fall, and limited on-site deployment in photovoltaic power plant construction. YOLOv8 was used as the baseline and the traditional C2f module was replaced with an improved dual-branch convolution fusion module (C2f-Dualconv) in order to enhance feature extraction and computational efficiency. The original neck network was replaced with a lightweight cross-scale feature fusion module (CCFM), introducing a multi-scale dilated Transformer attention (MSDA) between the backbone and the neck. The experiment was deployed on the Orange Pi5 Pro RK3588 platform. Results showed an average accuracy of 88.6%, with parameter count and computational load reduced by 31.3% and 22.2% respectively. Single-image detection time was 0.0297 s, and average power consumption was 2.2647 W. CMD-YOLO effectively addresses key challenges in fall detection at photovoltaic power plant through its advantage of low parameter count, low power consumption and high real-time performance. CMD-YOLO operates stably in resource-constrained field environment, providing reliable support for real-time detection on mobile device.