改进的YOLOv8n在复杂环境下的车辆识别算法

Improved YOLOv8n vehicle recognition algorithm under complex circumstances

【目的】针对城市复杂环境下的车辆难识别问题,提出了基于YOLOv8n(you only look once version 8n)的改进模型DB-YOLOv8n(deformable block YOLOv8n)。【方法】首先在颈部网络融合通道注意力机制(efficient channel attention,ECA)和改进加权双向特征金字塔网络(bidirectional feature pyramid network,BiFPN),以增强在昏暗光线下的车辆检测性能及对多尺度图像的处理能力,特别是对远处或部分遮挡的车辆;其次在主干网络引入可变型卷积(deformable convolutional networks,DCN),以增强模型对不同尺寸车辆的适应性;最后使用精确边界框回归的高效交并比损失函数(focal and efficient intersection over union loss,Focal-EIOU loss)替换高效交并比(efficient intersection over union,EIOU),进一步提升模型的稳定性。【结果】DB-YOLOv8n在自制车辆数据集上相比YOLOv8n,平均精度、精度和召回率分别提高了3.2%、3%和2%。【结论】本研究结果能为提高车辆检测的精确度提供理论参考。

[Objective]To address the difficulty of vehicle recognition under complex urban circumstances,an improved model named DB-YOLOv8n was proposed on the basis of YOLOv8n(you only look once version 8n).[Method]First,the neck network integrated the efficient channel attention(ECA)mechanism and an improved weighted bidirectional feature pyramid network(BiFPN)to enhance vehicle detection performance under low-light conditions and improve the handling of multi-scale images,particularly for distant or partially occluded vehicles;second,thebackbone network introduced the deformable convolutional networks(DCN)to enhance the model's adaptability to vehicles of different sizes;finally,the standard efficient intersection over union(EIOU)was replaced with the focal and efficient intersection over union loss(Focal-EIOU loss)to improve the model's stability and accuracy in bounding box regression.[Result]Compared to YOLOv8n,the proposed DB-YOLOv8n model improves the average precision,precision,and recall by 3.2%,3%,and 2%,respectively,on a custom vehicle dataset.[Conclusion]The results of this study can provide theoretical support for improving the accuracy of vehicle detection under complex circumstances.