【摘要】 语义分割技术对自动驾驶等实际场景的图像处理十分重要,然而基于传统相机的语义分割仍面临动态光照下信息缺失以及高速运动目标的运动模糊等问题.为此,引入高动态范围和高响应速度的事件相机,能够在恶劣光照和高速运动条件下有效成像,并提出了一种基于事件相机的图像语义分割(event-based image semantic segmentation, EVISS)方法.在数据集部分,针对基于事件相机的语义分割数据集较少且标注质量不高的问题,通过仿真环境制作了一个大规模和高精度标注的数据集Carla-Semantic;在网络设计部分,针对分布不均的事件数据特征难提取的问题, EVISS方法通过改进的图拉普拉斯公式和注意力机制,增强图像的全局联系性和上下文依赖,有效地提取高层级事件特征.在所制作的Carla-Semantic数据集上与现有技术Ev-SegNet进行实验的结果表明, EVISS方法在MPA和mIoU评价指标上分别达到87.89%和81.93%,超越了对比方法的表现,有效地实现基于事件相机的图像语义分割.更多还原

【Abstract】 Semantic segmentation technology is crucial for image processing in practical scenarios such as autonomous driving. However, semantic segmentation based on traditional cameras still faces challenges such as information loss under dynamic lighting and motion blur of high-speed moving objects. Therefore, this paper introduces event cameras with high dynamic range and high response speed, which can effectively image under harsh lighting and high-speed motion conditions, and proposes an Event-based Image Semantic Segmentation(EVISS) method based on event cameras. In the dataset section, a large-scale and high-precision annotated dataset, Carla-Semantic, was created through CARLA simulation environment to address the issue of a limited number of event camera based semantic segmentation datasets and low annotation quality. In the network design section, to address the problem of difficult feature extraction from event data with uneven distribution, the EVISS method enhances the global connectivity and context dependency of images through an improved graph Laplace formula and attention mechanism, effectively extracting high-level event features. The experimental results on the Carla-Semantic dataset and the existing method Ev-SegNet show that our EVISS method achieves 87.89% and 81.93% in MPA and mIoU evaluation metrics, respectively, surpassing the performance of the comparative methods and effectively achieving image semantic segmentation based on event cameras.更多还原