Abstract: In response to the challenges faced by machine guide dogs in detecting small targets such as traffic lights and unclear edge features like zebra crossings during navigation, as well as issues of missed detection, false detection, and repeated detection caused by varying backgrounds, this paper proposes an improved YOLOv8-based method for detecting typical obstacles and traffic signs for blind travelers. The method introduces the Triplet Attention mechanism into the backbone network of YOLOv8, significantly enhancing the model's ability to recognize key targets in complex environments by strengthening the capture of temporal relationships in sequences and reinforcing the correlation of local regions. Additionally, the algorithm integrates the CSFF (Cross Scale Feature Fusion) module and the TFE (Triple Feature Encoding) module, combining local and global feature information to obtain more precise feature representations, thereby improving detection accuracy.To further enhance the detection capability for small targets, the algorithm incorporates a small target detection head, P2, which assists in identifying small targets such as traffic lights by extracting lower-resolution features, thereby enhancing the model's multi-scale detection ability. At the same time, the Focaler-WIoU loss function is adopted to effectively reduce harmful gradients generated by samples, further optimizing the model's training process. Finally, feature distillation technology is employed to improve the accuracy of the enhanced algorithm, ensuring the model's efficiency and robustness in practical applications.Experimental results demonstrate that the improved algorithm achieves an average accuracy of 93.5% and a precision rate of 92.8% while reducing the number of parameters by 17.2%. This improvement not only significantly enhances the target detection capability of machine guide dogs in complex environments but also provides more accurate detection information for subsequent research on local path planning, holding significant practical value.

Key words: deep learning, obstacle detection, traffic signs detection, attention mechanism, knowledge distillation

摘要： 针对机器导盲犬在行进过程中面临的红绿灯等小目标和斑马线等边缘特征不清晰样本目标的检测难题，以及背景多变导致的漏检、错检和重复检测问题，本文提出一种改进的YOLOv8盲人出行的典型障碍物及交通标识检测方法。在YOLOv8的主干网络中引入Triplet Attention注意力机制，通过加强捕捉序列中的时序关系和强化局部区域的相关性，显著提升模型对复杂环境中关键目标的识别能力。此外，算法融合CSFF（Cross Scale Feature Fusion）模块和TFE（Triple Feature Encoding）模块，结合局部和全局特征信息，获得更加精确的特征表示，从而提高检测的准确性。为了进一步提升对小目标的检测能力，算法结合小目标检测头P2，通过提取较低分辨率的特征来帮助识别红绿灯等小目标，增强模型的多尺度检测能力。同时，采用Focaler-WIoU损失函数，有效减小样本产生的有害梯度，进一步优化模型的训练过程。最后，通过特征蒸馏技术对改进后的算法进行精度提升，确保模型在实际应用中的高效性和鲁棒性。实验结果表明，改进后的算法在参数量减少17.2%的情况下，平均准确率达到了93.5%，查准率为92.8%。这一改进不仅显著提升机器导盲犬在复杂环境中的目标检测能力，还为后续的局部路径规划研究提供更为准确的检测信息，具有重要的实际应用价值。

关键词: 深度学习, 障碍物检测, 交通标识检测, 注意力机制, 知识蒸馏