A Semantic SLAM Integrated with Enhanced YOLOv7 Target Detection Algorithm

This paper proposes a semantic SLAM integrated with an enhanced YOLOv7 target detection algorithm. To address the issue of image blurring caused by robot movement and camera shake, we have incorporated an image enhancement module before the tracking thread. Consequently, the resulting images are more clearer. In the feature extraction stage, we introduce adaptive thresholds to improve the system's capability in feature point extraction. To minimize the influence of dynamic objects on this system, we employ an enhanced YOLOv7 algorithm to detect dynamic targets. Then, we integrate it with epipolar constraint to eliminate dynamic feature points. Finally, We evaluated our system with five sequences taken from the TUM dataset, and compared with ORB-SLAM3, our system improves more than 91% in accuracy, up to 98%. Moreover, compared to similar semantic SLAM systems, our system offers improved accuracy as well as enhanced real-time performance.