Journal
FRONTIERS IN NEUROROBOTICS
Volume 16, Issue -, Pages -Publisher
FRONTIERS MEDIA SA
DOI: 10.3389/fnbot.2022.990453
Keywords
dynamic target detection; VSLAM; YOLOv3; GMM; Kalman filter
Funding
- Key Ramp
- D Program of Jiangsu Province (Industry Prospects and Key Core Technologies) [BE2020006-2]
- National Natural Science Foundation of China [61773219, 62003169]
- Natural Science Foundation of Jiangsu Province [BK20200823]
- Jiangsu Innovation and Entrepreneurship Talent Program Project [JSSCBS202030576]
- Natural Science Research Project of Jiangsu Higher Education Institutions [20KJB520029]
Ask authors/readers for more resources
This study proposed a real-time tracking and mapping method based on GMM combined with YOLOv3. By improving the tracking thread, using the Kalman filter algorithm, and filtering feature points, it achieves more stable augmented reality registration in dynamic environments.
Augmented Reality Registration field now requires improved SLAM systems to adapt to more complex and highly dynamic environments. The commonly used VSLAM algorithm has problems such as excessive pose estimation errors and easy loss of camera tracking in dynamic scenes. To solve these problems, we propose a real-time tracking and mapping method based on GMM combined with YOLOv3. The method utilizes the ORB-SLAM2 system framework and improves its tracking thread. It combines the affine transformation matrix to correct the front and back frames, and employs GMM to model the background image and segment the foreground dynamic region. Then, the obtained dynamic region is sent to the YOLO detector to find the possible dynamic target. It uses the improved Kalman filter algorithm to predict and track the detected dynamic objects in the tracking stage. Before building a map, the method filters the feature points detected in the current frame and eliminates dynamic feature points. Finally, we validate the proposed method using the TUM dataset and conduct real-time Augmented Reality Registration experiments in a dynamic environment. The results show that the method proposed in this paper is more robust under dynamic datasets and can register virtual objects stably and in real time.
Authors
I am an author on this paper
Click your name to claim this paper and add it to your profile.
Reviews
Recommended
No Data Available