基于环视相机的无人驾驶汽车实例分割方法

Surround view cameras based instance segmentation method for autonomous vehicles

针对环视鱼眼图像中目标几何畸变大导致建模难的问题,提出一种基于可变形卷积网络的实例分割方法,主要是在Mask R-CNN框架的基础上引入可变形卷积和可变形RoI Pooling(候选区域池化)来提升网络对几何畸变的建模能力.针对深度神经网络训练数据缺乏、易过拟合的问题,提出了基于多任务学习的训练方法.首先将现有的大规模普通图像数据集转换为鱼眼数据集来弥补训练数据不足的问题,然后采用多任务学习的训练方法将转换的图像和真实图像放在同一个框架中训练以提高网络的泛化能力.用该方法在真实的环视鱼眼图像上做测试,结果表明:相对于原始Mask R-CNN的方法平均精度提升了3.1%,证明了该方法在真实交通环境中的有效性.

Aimed at handling the problem of modeling large geometric distortions caused by surround view cameras,a deformable convolutional networks based instance segmentation method was proposed.The method introduced the deformable convolution and deformable RoI(region of interest) Pooling into the framework of Mask R-CNN.Aimed at handling the problems of insufficient data to train the deep neural networks and overfitting, a multi-task learning based training method was proposed. First, an existing large-scale dataset of conventional images was transformed to a fisheye-style dataset to compensate the lack,and then a multi-task learning method was adopted to train the transformed images and real-world images in a united architecture to improve the generalization ability.The proposed method was tested on the real-world fisheye images.It shows an improvement of 3.1% over the original Mask R-CNN method,which demonstrates the effectiveness of the proposed method in real-world traffic environments.