复杂气象条件下的交通场景目标检测算法研究

Study on Traffic Scene Object Detection Algorithm under Complex Meteorological Conditions

针对复杂条件下交通场景目标检测虚警和漏检问题,提出一种基于YOLOv3的快速检测算法。首先利用kmeans算法聚类出适合该数据集的合理的anchors坐标;其次在YOLOv3框架基础上使用密集模块代替残差网络,加强特征的传播和复用,并进行多尺度融合;将普通卷积替换为空洞卷积,在不改变网络层数和计算量的基础上增大感受野。另外针对质量不佳的图片利用暗通道去雾算法对图片进行增强处理。经过试验证明,经过改进的YOLOv3算法在数据集上准确率和召回率均有明显提升,具有很强的通用性和鲁棒性,且参数数量明显减少。

Aiming at the problem of false alarm and missed detection of traffic scene targets under complex conditions,a fast detection algorithm based on YOLOv3 is proposed.Firstly,the kmeans algorithm was used to cluster the reasonable anchors coordinates suitable for the data set.Secondly,the dense module was used instead of the residual network based on the YOLOv3 framework to enhance the feature propagation and reuse,and conduct multi-scale fusion.The ordinary convolution was replaced with the cavity convolution,and the receptive field was increased without changing the number of layers and the amount of calculation.In addition,the dark channel dehazing algorithm was used to enhance the image for poor quality images.The experimental results show that the improved YOLOv3 algorithm has improved the accuracy and recall rate in the dataset of this paper.It has strong versatility and robustness,and the number of parameters is significantly reduced.