基于cnn卷积神经网络的特征点提取与相机估计研究

Study on Feature Point Extraction and Camera Estimation Based on CNN Convolution Neural Network

该文使用有两个深卷积神经网络的点追踪系统。第一个网络,Magic Point,对单个图像进行操作,并提取显著的2D特征点。提取的点是"为SLAM准备好的",因为这些点是孤立设计的并且在整个图像中分布良好。我们将该网络与经典点检测器进行比较,并发现图像存在噪声的情况下特征点提取性能有显著的改进。当检测到的点几何稳定时,变换估计更简单,我们设计了一个第二个网络Magic Warp,它可以对成对的点图像(Magic Point的输出)进行操作,并估计与输入相关的单应性。这种转换引擎与传统的方法不同,因为它不使用本地点描述符,只是点位置。这两个网络都用简单的合成数据进行训练,减轻了高代价的外部摄像机地面实况和高级图形渲染的需求。该系统快速,精简,在单个CPU上轻松运行30+FPS。

In this paper,we use a point tracking system with two deep convolution neural networks.The first network,MagicPoint,operates on a single image and extracts significant 2D feature points.The extracted point is ＂ready for SLAM＂,because these points are designed in isolation and are well distributed throughout the image. We compare the network with the classical point detector,and find that the feature point extraction performance is greatly improved when the image is noisy.When the detected point geometry is stable,the transformation estimation is simpler.We design a network of second MagicWarp,which can operate pairs of point images （MagicPoint output） and estimate homography related to input.This transformation engine is different from the traditional method,because it does not use the local point descriptor,but it is just a point position.These two networks are trained with simple synthetic data to reduce the demand for high cost external camera ground live and advanced graphics rendering.The system is fast,streamlined,and easily runs 30＋ FPS on a single CPU.