【摘要】 针对城市道路环境复杂、草坪树木密集等引起的树木误检、漏检以及模型较大不易部署在边缘计算设备上等问题，提出一种基于改进YOLO v5的城市树木检测算法.算法引入融合注意力机制的双向加权特征金字塔结构BiFPN-ECA,减少冗余语义特征，增加特征融合网络对关键特征的关注度，提升特征融合质量；利用GSConv卷积替换Neck部分的密集卷积，提升特征提取能力并减少运算量；引入SIoU损失函数，通过加入向量角度这一惩罚项，减少与距离相关的变量，降低回归自由度，加快网络收敛，进一步提升回归精度.采集城市常见绿化树木的图像和视频，建立城市绿化树木样本集进行模型对比试验.试验结果表明，本文方法比初始YOLO v5模型的平均精度均值提升了5%,模型参数量减小为原来的50%,模型的帧率提升了5帧/s,兼顾了检测精度和实时性.更多还原

【Abstract】 Addressing the challenges of false detection and missed detection of trees due to complex urban road environment, dense lawns and trees, along with the difficulty of deploying the model on embedded devices, this paper proposes an urban tree detection algorithm based on an enhanced YOLO v5. The algorithm introduces a bidirectional weighted feature pyramid network with efficient channel attention(BiFPN-ECA) that incorporates a fusion attention mechanism to reduce redundant semantic features, enhance the network’s focus on key features, and improve the quality of feature fusion. By utilizing GSConv convolution to replace the dense convolution in the Neck part, the algorithm enhances feature extraction capability while reducing computational load. The algorithm adopts the SIoU loss function, incorporating a penalty term for vector angle, which reduces variables related to distance, lowers the regression degree of freedom, accelerates network convergence, and further improves regression accuracy. Through collecting images and videos of common urban greeneries, and setting up a sample set of urban greeneries for model comparative experiments, the experimental results show that the proposed method improves the average precision by 5% compared to the original YOLO v5 model, reduces the model parameter size to 50%, and increases the model’s frame rate by 5 frames/s, achieving a balance between detection accuracy and real-time performance.更多还原