基于改进YOLOv3算法的公路车道线检测方法

Road Lane Line Detection Method Based on Improved YOLOv3 Algorithm

针对YOLOv3算法在检测公路车道线时存在准确率低和漏检概率高的问题,提出一种改进YOLOv3网络结构的公路车道线检测方法.该方法首先将图像划分为多个网格,利用K-means++聚类算法,根据公路车道线宽高固有特点,确定目标先验框数量和对应宽高值;其次根据聚类结果优化网络Anchor参数,使训练网络在车道线检测方面具有一定的针对性;最后将经过Darknet-53网络提取的特征进行拼接,改进YOLOv3算法卷积层结构,使用GPU进行多尺度训练得到最优的权重模型,从而对图像中的车道线目标进行检测,并选取置信度最高的边界框进行标记.使用Caltech Lanes数据库中的图像信息进行对比试验,实验结果表明,改进的YOLOv3算法在公路车道线检测中平均准确率(Mean average precision, mAP)为95%,检测速度可达50帧/s,较YOLOv3原始算法mAP值提升了11%,且明显高于其他车道线检测方法.

Aiming at the problem that the YOLOv3 algorithm has low accuracy, high probability of missed detection when detecting road lane lines, a road lane detection method for improving YOLOv3 network structure is proposed. At first, the method divides the image into multiple grids, and uses the K-means++ clustering algorithm to determine the number of target priori boxes and the corresponding value according to the inherent characteristics of the road lane line width and height. Then, according to the clustering result, the network anchor parameter is optimized to make the training network have certain pertinence in lane line detection. At last, the features extracted by the Darknet-53 are spliced, the network structure of the YOLOv3 algorithm is improved, and the GPU is used for multi-scale training to obtain the optimal weight model, thereby detecting the lane line target in the image and selecting the bounding box with the highest confidence to mark. Using the image information in the Caltech Lanes database for comparison experiments, the experimental results show that the improved YOLOv3 algorithm’s mean average precision is 95% in road lane detection, the improved detection speed can be achieved 50 frame/s, which is11% higher than the original algorithm and significantly higher than other lane detection methods.