| Summary: | Abstract Rockfalls on mountainous roads pose significant safety risks to pedestrians and vehicles, particularly in remote areas with underdeveloped communication infrastructure. To enable efficient detection, this study proposes a rockfall detection system based on embedded technology and an improved Yolov8 algorithm, termed Yolov8-GCB. The algorithm enhances detection performance through the following optimizations: (1) integrating a lightweight DeepLabv3+ road segmentation module at the input stage to generate mask images, which effectively exclude non-road regions from interference; (2) replacing Conv convolution units in the backbone network with Ghost convolution units, significantly reducing model parameters and computational cost while improving inference speed; (3) introducing the CPCA (Channel Priori Convolution Attention) mechanism to strengthen the feature extraction capability for targets with diverse shapes; and (4) incorporating skip connections and weighted fusion in the Neck feature extraction network to enhance multi-scale object detection. Experimental results demonstrate that Yolov8-GCB improves AP@0.5 and AP@0.75 by 1.2% and 1%, respectively, while reducing the number of parameters by 14.1% and the GFLOPs by 16.1% and increasing inference speed by 20.65%. This method provides an effective technological solution for real-time rockfall detection on embedded devices and can be extended to other disaster scenarios, such as landslides and debris flows, in regions with limited infrastructure.
|
|---|
