| 基于YOLOv8 的道路交通目标检测优化算法 |
| 点此下载全文 |
| 引用本文：付俊强,谭金菁,高瑞玲,谭文安,毛文逸.基于YOLOv8 的道路交通目标检测优化算法[J].上海第二工业大学(中文版),2025,42(4):433-440 |
| 摘要点击次数: 193 |
| 全文下载次数: 22 |
|
| 基金项目:国家自然科学基金项目(61672022, U1904186), 上海市教委青年教师培养项目(A60GY24C002) 资助 |
|
| 中文摘要:当今交通目标检测中, 小目标检测仍存在一定漏检率, 且在复杂道路背景下目标识别率较低等缺点。针对上述问题, 本文提出了一种基于YOLOv8 的道路交通目标检测优化算法。具体针对小目标检测漏检情况, 引入高效多尺度注意力(efficient multi-scale attention, EMA) 机制, 以增强模型对小目标的感知能力; 使用第3代可变形卷积网络(deformable convolutional networks version 3, DCNv3) 替代模型中普通卷积, 通过动态调整采样位置提升特征图中不规则形状的特征提取能力, 从而改善模型对遮挡和重叠目标的检测效果。同时采用基于最小封闭框的交并比(generalized intersection over union, GIoU) 损失函数, 进一步优化交通目标检测算法。改进算法在测试数据集中进行实验,平均精度均值mAP50 与mAP50:95 分别达到87.4% 和53.6%, 对比基线模型分别提升了5.2% 和3.7%, 有效验证了所提优化模型在复杂道路场景下进行交通目标检测的可行性。 |
| 中文关键词:YOLOv8 注意力机制 可形变卷积 复杂道路场景 |
| |
| Road Traffic Target Detection Optimization Algorithm Based on YOLOv8 |
|
|
| Abstract:Nowadays, in traffic target detection, small target detection still has a certain missed detection rate, and the target recognition rate is low in complex road backgrounds. In response to the above problem, an optimized algorithm for road traffic target detection based on YOLOv8 is proposed. Specifically, for the missed detection of small targets, an efficient multi-scale attention (EMA) mechanism is introduced to improve the model’s perception ability of small targets. Deformable convolutional networks version 3 (DCNv3) is used to replace the ordinary convolution in the model. By dynamically adjusting the sampling positions, the ability to extract features of irregular shapes in the feature map is enhanced, thus improving the model’s detection performance for occluded and overlapping targets. At the same time, a generalized intersection over union (GIoU) loss function based on the minimum bounding box is employed to optimize the traffic target detection algorithm. The improved algorithm is experimented on the test datasets. The mean average precision mAP50 and mAP50:95 reach 87.4% and 53.6%, which are increased by 5.2% and 3.7% respectively compared with the baseline model. This effectively proves the proposed optimized model is feasible for traffic target detection in complex road scenarios. |
| keywords:YOLOv8 attention mechanism deformable convolution complex road scene |
| 查看全文 查看/发表评论 下载PDF阅读器 |
|
|
|