| 摘要: |
| 目的 视觉 SLAM 作为自动驾驶和移动机器人的核心技术之一,传统算法无法应对高度动态的环境,也缺乏
地图的语义信息,解决动态物体对 SLAM 系统的影响是研究的主要目标,也是当前热点问题之一。 方法 提出一个
新的基于 YOLOv8 同步动态检测与局部语义分割的方法,来实现动态环境下的位姿估计与局部语义建图。 首先,
通过应用 YOLOv8 对输入图像进行同步动态检测和语义分割,使用目标检测结果的目标框对动态特征点进行剔
除,再运用静态特征点进行姿态估计,然后在系统的语义建图线程中,对语义分割后的图像加入扩张掩模,最后使
用点云库进行语义地图的构建,从而产生能够应用于实际场景的语义地图。 结果 在 TUM 数据集中进行了比较试
验,数据显示:这种方法相对于传统方法能提高 98. 1%的位姿准确率,并且在实时性测试中,本文算法的速度也优
于同类算法,而且可以在同一时间创建出局部语义地图。 结论 基于 YOLOv8 同步动态检测与局部语义的方法来处
理常规场景下的动态物体对 SLAM 系统的影响十分有效,且实时性高,但对于一些特殊场景如摄像机大幅旋转等,
由于目标检测的失效而导致动态特征剔除失败,从而系统精度降低。 |
| 关键词: 视觉 SLAM 语义分割 目标检测 语义地图 |
| DOI： |
| 分类号: |
| 基金项目: |
|
| Synhronous Dynamic Detection and Local Semantic Visual SLAM Based on YOLOv8 |
|
YANG Haibo CAO Chuqing
|
|
School of Computer and Information Anhui Polytechnic University Anhui Wuhu 241000 China
|
| Abstract: |
| Objective Visual SLAM one of the core technologies for autonomous driving and mobile robots is currently
unable to cope with highly dynamic environments with traditional algorithms and lacks semantic information in maps.
Addressing the impact of dynamic objects on SLAM systems is the main objective of this study which is also one of the
current hot topics. Methods A novel method based on YOLOv8 synchronous dynamic detection and local semantic
segmentation was proposed to realize the position estimation and local semantic map building in dynamic environments.
Firstly synchronous dynamic detection and semantic segmentation were performed on input images using YOLOv8.
Dynamic feature points were eliminated using the target frame of the target detection result and then static feature points
were applied for pose estimation. Then an expansion mask was added to the semantically segmented image in the
semantic mapping thread of the system. Finally a point cloud library was used for the construction of semantic maps
thereby generating semantic maps applicable to real-world scenarios. Results Comparative tests were conducted in the
TUM dataset and the data showed that this method can improve the position accuracy by 98. 1% compared with traditional
methods. Moreover the speed of the proposed algorithm was superior in real-time testing compared with similar algorithms
and it can create local semantic maps simultaneously. Conclusion The method based on YOLOv8 for synchronous dynamic detection and local semantics is highly effective in addressing the impact of dynamic objects on SLAM systems in typical
scenes with high real-time performance. However in some special scenarios such as significant camera rotation the failure
of object detection leads to the failure of dynamic feature removal resulting in reduced system accuracy. |
| Key words: visual SLAM semantic segmentation target detection semantic maps |
