Optimized Lightweight YOLOv5 Model for Efficient Wildfire Flame and Smoke Detection on Edge Computing Devices

Although unmanned aerial vehicles (UAVs) equipped with deep learning–based object detection show promise for early wildfire detection, their performance is often constrained by edge computing devices that lack graphic processing unit (GPU) acceleration. This research addresses these limitations by introducing a novel one-stage object detection algorithm based on the You Only Look Once version 5 nano (YOLOv5n) architecture, optimized specifically for central processing unit (CPU)-based edge devices. Key enhancements include modifying the backbone and neck to improve detection speed, power consumption, and overall energy efficiency. MobileNetV3-Small is employed as a lightweight backbone, while Ghost Convolution modules are incorporated in the neck to reduce complexity and boost inference speed on CPUs. Additional simplifications come from halving the number of output channels in the neck. To further elevate accuracy, the Bottleneck module is adjusted to use 3×3 kernels instead of 1×1 kernels in the first convolution. Finally, the trained PyTorch weights are exported to two edge-specific frameworks, Open Neural Network Exchange (ONNX) and Open Visual Inference and Neural Network Optimization (OpenVINO), to accelerate CPU-based inference. Compared to the original YOLOv5n, the modified model converted to OpenVINO demonstrates a 423% increase in inference speed—up to 31.9 frames per second (FPS)—along with a 11.4% reduction in power consumption on a CPU-based edge device. Overall, these findings highlight the optimized algorithm’s potential for real-time, early wildfire detection.