基于改进YOLOv8s的自动驾驶多目标跟踪检测研究
Research on Improved Multi-Target Tracking Detection of YOLOv8s in Autonomous Driving Scenario
王轩慧 1吴颖 1邵凯扬 2谢德燕 1董建业3
作者信息
- 1. 青岛农业大学,理学与信息科学学院,青岛 266109
- 2. 青岛农业大学,机电工程学院,青岛 266109
- 3. 中国电子科技集团公司第二十二研究所,青岛 266109
- 折叠
摘要
针对现有自动驾驶模型对小样本及重叠样本识别精度不高的问题,提出了一种基于改进YOLOv8s的轻量级目标检测模型.使用多尺度特征提取设计了C2f-Faster模块,替换YOLOv8s骨干网络与颈部网络的C2f模块;融合内部交并比(Inner-IoU)与基于最小点距离交并比(MPDIoU)损失函数,提出Inner-MPDIoU损失函数.模型的对比试验、消融试验结果表明:交并比为0.5时模型平均精度(mAP@0.5)提升3.5百分点,准确率达到95.2%,参数量下降25%.通过数据的可视化分析,进一步验证了改进模型对于复杂场景的有效性.
Abstract
In order to solve the problem that the existing autonomous driving models do not have high recognition accuracy for small samples and overlapping samples,a lightweight object detection model based on improved YOLOv8s is proposed.A C2f-Faster module is designed by using multi-scale feature extraction to replace the C2f module of YOLOv8s backbone network and neck network.The Inner-MPDIoU loss function is proposed by fusing the Inner-IoU and MPDIoU-based loss function based on the Minimum Point Distance based Intersection over Union(MPDIoU).The results of the comparative test and ablation experiment of the model show that when the cross-union ratio is 0.5,the average accuracy of the model(mAP50)is increased by 3.5 percentage points,the accuracy reaches 95.2%,and the number of parameters decreases by 25%.Through the visual analysis of the data,the effectiveness of the improved model for complex scenarios is further verified.
关键词
自动驾驶/深度学习/目标检测/YOLOv8s/损失函数Key words
Autonomous driving/Deep learning/Object detection/YOLOv8s/Loss function引用本文复制引用
出版年
2024
NETL
NSTL
万方数据