基于雾天条件下交通道路上的目标检测

Object Detection of Traffic Road Based on Fog

在目前的道路车辆的检测当中,对车辆的检测精度和匹配度的要求越来越高,尤其在雾天的天气情况下会对车辆的检测产生影响。针对以上问题,深受深度学习理论的启发,提出一种在去雾网络与目标检测网络结合情况下,对雾天下交通道路上车辆图像进行去雾目标检测的深度学习目标检测算法,并且模型轻量,易于嵌套使用。实验结果表明将YOLOv5目标检测网络与AOD-Net去雾相结合,在真实和合成的有雾的数据集上,该算法在定量评价和定性评价中均优于对比算法。其中MAP值达到81.73%,比YOLOv5算法的MAP值高1.31%且FPS达到了25.0,速度得到提升,表明AOD-Net与YOLOv5网络相结合的算法能更加有效地检测雾天条件下道路交通的车辆,且网络泛化能力和鲁棒性较好。

In the current detection of road vehicles, the detection accuracy and matching degree of vehicles are increasingly high,which affect the detection of vehicles especially in foggy weather conditions. In view of the above problems, deeply inspired by Deep Learning theory, a Deep Learning target detection algorithm is proposed to detect defogging targets in vehicle images on traffic roads in foggy days under the combination of defogging network and target detection network. The model is lightweight and easy to be nested. Experimental results show that combining YOLOv5 target detection network with AOD-Net defogging, the algorithm is superior to the comparison algorithm in both quantitative and qualitative evaluation on real and synthetic foggy data sets. The MAP value reaches 81.73%, 1.31% higher than that of YOLOv5 algorithm, and the FPS reaches 25.0, which improves the speed. The results show that the algorithm combined with AOD-NET and YOLOv5 network can detect vehicles in foggy road traffic more effectively, and the network generalization ability and robustness are better.