无人驾驶车辆基于语义分割方法障碍物检测

Obstacle detection based on semantic segmentation for unmanned vehicles

障碍物检测是无人驾驶车辆环境感知重要的组成部分,语义分割技术能够对障碍物进行像素级检测。为满足无人车系统的实时性要求和对障碍物检测精度要求,提出了一种轻量级语义分割模型。该模型构建了特征提取块,通过跳跃层结构将底层级特征与高层级特征相融合,用于提取更加细化的图像特征信息。运用深度可分离卷积代替标准卷积操作,减少了模型参数量和计算量。利用不同膨胀率的膨胀卷积以获取多尺度目标信息,在上采样时融合不同尺度的特征信息,使得语义信息更加丰富。试验结果表明:提出的轻量级语义分割模型在Cityscapes数据集和ApolloScape数据集上取得了较好的障碍物检测结果,同时也满足无人车的实时性要求。

Obstacle detection is an important part of the environment perception of unmanned vehicles.Semantic segmentation technology can detect the obstacles at the pixel level.In order to meet the real-time requirements of unmanned vehicle systems and the requirements for obstacle detection accuracy,a lightweight semantic segmentation model was proposed.In this paper,a feature extraction block was constructed,which combined the low-level features with the high-level features through the skip layer to extract more refined image feature information.The use of depthwise separable convolution instead of standard convolution operations reduced the amount of model parameters and computation cost.The expansion convolution with different expansion rates was used to obtain multi-scale target information.The feature information of different scales was fused at the time of upsampling,so that the semantic information was more abundant.The proposed model achieves great segmentation results in the Cityscapes dataset and ApolloScape dataset,and also met the real-time requirements of unmanned vehicles.