Abstract: On normally operating highways, there are dangerous targets that interfere with drivers’ judgment and pose traffic hazards. When using drones for detection, there are challenges such as occlusion, overlap, dispersion and heterogeneity. To address these issues, a high-precision detection algorithm based on YOLOv8n, called CT-YOLO, is proposed. Firstly, dilated convolution is reconstructed in the C2f module of the YOLOv8 backbone network, and 1×1 convolutions are integrated before and after the convolution to solve the problem of decentralized targeting of application scenarios. Secondly, the classic feature pyramid network is improved, and two additional detection layers are added to enhance detection accuracy for occluded targets. Lastly, an improved triple attention mechanism is integrated into the Head part of the C2f module to enhance the model’s ability to capture heterogeneous target information. An image dataset containing 11 types of anomalous targets, including fractures, patches, pericarp, leaves, plastic, potholes, arrows, lane lines, cardboard boxes, oil, and cans, is constructed through video collection, frame extraction, manual annotation, and data augmentation. Experimental results indicate that the CT-YOLO algorithm improves mAP@0.5 by 13.2 percentage points and mAP@0.5：0.95 by 11 percentage points on the anomalous target image dataset, significantly enhancing detection accuracy and demonstrating good practical application effectiveness.

Key words: highway, unmanned aerial vehicle(UAV), YOLOv8, target detection, multiple targets, small targets

摘要： 正常运行的高速公路上，存在干扰驾驶员判断、造成交通隐患的危险目标，使用无人机进行检测时可能面临遮挡、重叠、分散、异构等难点。为解决这些问题，提出一种基于YOLOv8n的高精度检测算法——CT-YOLO。在YOLOv8骨干网络C2f模块中重构空洞卷积（dilated convolution），在卷积前后分别融合1×1卷积，解决应用场景目标分散的问题；改进经典特征金字塔网络，额外增加两个检测层，提高了对遮挡、小目标的检测精度；将改进的三重注意力机制融合到Head部分的C2f模块中，增强模型对异构目标信息的捕捉能力。通过视频采集、分帧、人工标注和数据增强，构建了一个包含11种异常目标的图像数据集，包括裂缝、修补、果皮、树叶、塑料、坑槽、箭头、车道线、纸箱、泛油和易拉罐。实验结果表明，CT-YOLO算法在异常目标图像数据集上mAP@0.5提升了13.2个百分点，mAP@0.5：0.95提升了11个百分点，检测精度明显提高，具有较好的实际应用效果。

关键词: 高速公路, 无人机（UAV）, YOLOv8, 目标检测, 多目标, 小目标