摘要
在图像清晰、无遮挡的情况下,车辆检测模型可以准确识别出大多数车辆,但在雾霾和雨雪等极端天气下,特征和输出标签的联合分布改变,准确率大幅下降。为平衡速度与准确度,本文提出单阶段目标检测网络与域适应结合的JMMD-YOLO模型,对齐跨域的多个特定层的联合分布,使模型同时学习雾霾特征与晴天特征,改进后的模型指标有所提升,在150 m可见度下单类平均精度最高提升6.4%,整体平均精度提升2.2%。改进后的模型在雾天情况下能更好地识别出车辆目标,为复杂条件下的车辆目标检测或目标跟踪提供参考。
With clear and unobstructed images,the vehicle detection model can accurately identify most vehicles. However,in extreme weather such as haze,rain,and snow,the joint distribution of features and output labels changes,and the accuracy drops significantly. In order to balance speed and accuracy,this paper proposes a JMMD-YOLO model that combines a single-stage object detection network with domain adaptation to align the joint distribution of multiple specific layers across domains,so that the model can learn features of haze and sunny day at the same time. The indicators of the improved model have been improved. Under the visibility of 150 meters,the average accuracy of a single class is increased by up to 6.4%,and the overall average accuracy is increased by 2.2%. The improved model can better identify vehicle targets in foggy conditions,and provide a reference for vehicle target detection or target tracking under complex conditions.
作者
董惠文
田莹
DONG Huiwen;TIAN Ying(School of Computer Science and Software Engineering,University of Science and Technology Liaoning,Anshan,114051,China)
出处
《辽宁科技大学学报》
CAS
2022年第2期104-112,共9页
Journal of University of Science and Technology Liaoning
基金
辽宁省教育厅项目(2019LNJC03)。
关键词
车辆目标检测
雾天目标检测
域适应
vehicle object detection
target detection in fog
domain adaptation