Object Semantic Annotation Based on Visual SLAM

Abstract

In order to solve the problem that mobile robots cannot obtain the environment map with semantic information in robot localization and map reconstruction, the classical SLAM algorithm ORB-SLAM2 and target detection algorithm YOLO-v3 are studied and improved, and an improved YOL-SLAM algorithm is proposed. Firstly, sub thread is added in ORB-SLAM2. After selecting the key frame, the object detection and annotation are carried out on the key frame, and more semantic information is given to the map. At the same time, the YOLO-v3 network framework, darknet53, is simplified to reduce the number of original convolution layers, resulting in a lighter network structure, which is more in line with the real-time requirements of SLAM. The experimental results show that the optimized YOLO-v3 target detection rate increases by 217%, which meets the real-time requirements. The experimental platform based on Kinect V2 is built to test the performance of pal-SLAM algorithm. The results show that YOL-SLAM can also complete trajectory reconstruction and semantic annotation in indoor and outdoor environment, which not only meets the requirements of real-time, but also has good robustness.