改进YOLOV4-Tiny的雨雾道路环境下的实时目标检测

A Fast Target Detection Algorithm in Severe Road Environment Based on Improved YOLOV4-Tiny

为了提高恶劣道路场景下的目标检测能力,本文在YOLOV4-Tiny的基础上提出了一种快速目标检测算法。首先,本文考虑雨雾天气条件下的道路目标检测场景,将基于二次模糊的清晰度算法(ReBlur)和暗通道先验算法相结合对图像进行处理,然后将处理前后的图像用于网络的训练和测试,以克服雨雾天气造成的图像质量下降问题;另一方面,考虑道路场景中的小目标检测,本文在原网络的基础上对8倍降采样特征图进行上采样,再把得到的上采样结果与上一层的特征图拼接,以添加针对小目标的检测头。实验结果表明,改进后的网络在复杂道路场景下的检测能力显著提高,整体的平均精度均值(Mean Average Precision,mAP)也提高了4.13,同时检测速度达到了213 FPS。

In order to improve the algorithm’s detection ability in severe road scenes,a fast target detection algorithm based on YOLOV4-Tiny was proposed.First,in consideration of severe weather conditions,this paper combined the algorithm of ReBlur and the dark channel prior algorithm to process the images.On the foundation of the above results,the images before and after processing are used for network training and testing to overcome the problem of image quality degradation.On the other hand,for the detection of small targets,a detection head for small targets was added.The eight times down-sampling feature map was up-sampled for splicing with the feature map of the upper layer.The experimental results show that the improved network’s detection ability is significantly improved in complex road scenes,and the overall mean average precision(mAP)value is also increased by 4.13.Meanwhile,the detection speed reaches 213 FPS.