|
Simultaneous Localization and Mapping (SLAM) system is very important for autonomous driving. The problem of data association between features becomes the bottleneck limiting the performance of traditional visual SLAM systems, especially in complex environments. In computer vision tasks, Convolutional Neural Networks (CNN) have better performance in complex environments than traditional methods. Therefore, many studies combine SLAM systems with CNN for more reliable data association. In this paper, we design a CNN network for extracting descriptors and combine it with hand-crafted keypoints to construct a visual SLAM system. The experiments in this paper show that CNN-based local descriptors can significantly improve the accuracy and robustness of the SLAM systems. Compared with traditional visual SLAM systems, the SLAM system in this paper is more robust in complex environments, and the localization error of the system in this paper is 24.4% and 33.3% lower than ORB-SLAM2 and VINS-Mono on the evaluated datasets. Meanwhile, CNN local descriptors can be combined with any visual SLAM system, this method has good portability. |
|
Keywords:Computer Vision; Visual SLAM; CNN; Local Descriptors |
|