基于改进Mask R-CNN+LaneNet的车载图像车辆压线检测

Vehicle pressure line detection based on improved Mask R-CNN+LaneNet

针对辅助驾驶或自动驾驶领域车载图像的车辆压线检测问题,以及检测过程中由于欠曝、阴影或实体遮挡等因素导致的漏检、误检问题,提出基于改进Mask R-CNN与LaneNet的车辆压线检测算法。在网络优化方面,在Mask RCNN网络的基础上将RoI Align层的图像缩放算法(双线性插值)改进为双三次插值,将全连接层卷积化的VGG16网络取代LaneNet的E-Net共享解码器;在图像增强方面,改进Gamma校正算法以实现欠曝图像的自动校正;在训练数据方面,完成Tusimple数据集中车辆目标的标注并基于改进的随机擦除算法在网络训练过程中进行数据增强。实验结果表明:车辆检测速度保持不变的同时车道线检测速度提升了28%,车辆漏检率、误检率分别降低了38.93%,89.04%,车道线漏检率、误检率分别降低了67.21%,87.05%,算法的性能指标可满足车辆压线判断的应用需求。

To address the problem of vehicle pressure line detection of on-board images in the field of assisted or automatic driving,as well as the problem of missed and false detection caused by underexposure,shadow,or solid occlusion in the detection process,a vehicle pressure line detection algorithm based on improved Mask R-CNN and LaneNet was proposed.In terms of network optimization,based on the Mask R-CNN network,the image scaling algorithm(bilinear interpolation)of the ROI alignment layer was improved to bicubic interpolation,and the convoluted VGG16 network of a full connection layer was replaced by LaneNet’s E-Net shared decoder.For image enhancement,the Gamma correction algorithm was improved to realize the automatic correction of underexposed images.In terms of training data,the vehicle target in the Tusimple data set was marked,and the data were enhanced in the network training process,based on the improved random erasing algorithm.The experimental results show that while the vehicle detection speed remains unchanged,the lane line detection speed is increased by 28%,and the vehicle missed and false detection rate are reduced by 38.93%and 89.04%,respectively.Further,the lane line missed and false detection rate are reduced by 67.21%and 87.05%,respectively.The achieved performance index can meet the requirements of vehicle line pressing judgement method.