Research on Real-Time Semantic SLAM Based on Object Detection Network

Simultaneous localization and mapping (SLAM) is an important technology in the field of robotics. Semantic SLAM can provide a more accurate localization and satisfy the needs of complex applications, which has become a research hot spot. In this paper, we propose a real-time semantic SLAM system that uses semantic information in loop closure detection and mapping. In the system, we use object detection network to get semantic information including bounding box and category. In loop closure detection, we only use semantic information to construct feature structure, and implement feature comparison. Compared with the Bag of Visual Words (BoVW), the proposed approach does not need to generate vocabulary, holds very less amount of memory. Besides, we propose a fast semantic segmentation method combining bounding box and RGB-D image to create semantic OctoMap. Finally, we evaluate our semantic SLAM. Experimental results show that compared with BoVW, the proposed loop closure detection algorithm is about 20% higher in accuracy under the same recall rate. The semantic segmentation results reveal that the mean pixel accuracy of our method on NYUv2 dataset is 0.67.And our system takes 150ms to infer an image with the size of 640480 on MX150 GPU,i7-8550U CPU, which can be used in real-time visual SLAM.