# Dynamic SLAM by Combining Rigid Feature Point Set Modeling and YOLO

## TL;DR

This paper introduces a dynamic SLAM system that combines object recognition and depth segmentation to improve real-time tracking and positioning accuracy in dynamic environments.

## Contribution

The paper proposes adaptive rigid point set modeling and integrates Kalman filtering with DIoU for dynamic object tracking in SLAM.

## Key findings

- The system achieves real-time dynamic feature point screening.
- It effectively tracks occluded dynamic targets.
- Positioning accuracy in dynamic environments is improved.

## Abstract

What are the main findings?
This paper combines target recognition and depth threshold segmentation to rapidly segment the target point cloudThis paper integrates Kalman filtering and the depth intersection-over-union method (DIoU) for the association of target bounding boxes.This paper proposes the adaptive Rigid Point Set Modeling and creates rigid and non-rigid factors for the factor graph optimization in the SLAM system.

This paper combines target recognition and depth threshold segmentation to rapidly segment the target point cloud

This paper integrates Kalman filtering and the depth intersection-over-union method (DIoU) for the association of target bounding boxes.

This paper proposes the adaptive Rigid Point Set Modeling and creates rigid and non-rigid factors for the factor graph optimization in the SLAM system.

What are the implications of the main findings?
Achieve real-time screening of dynamic feature points.When occlusions occur among dynamic targets, the system can still effectively track the targets.Improve the positioning accuracy of the SLAM system in dynamic environments.

Achieve real-time screening of dynamic feature points.

When occlusions occur among dynamic targets, the system can still effectively track the targets.

Improve the positioning accuracy of the SLAM system in dynamic environments.

To obtain accurate location information in dynamic environments, we propose a dynamic visual–inertial SLAM algorithm that can operate in real-time. In this paper, we combine the YOLO-V5 algorithm and the depth threshold extraction algorithm to achieve real-time pixel-level segmentation of objects. Meanwhile, to address the situation where dynamic targets are occluded by other objects, we design the object depth extraction method based on K-means clustering. We also design a factor graph optimization with rigid and non-rigid dynamic objects based on object category division, in order to better utilize the motion information of dynamic objects. We use the Kalman filter algorithm to achieve object matching and tracking. At the same time, to obtain as many rigid targets as possible, we design the adaptive rigid point set modeling algorithm to further supplement the rigid objects. Finally, we evaluate the algorithm through public datasets and self-built datasets, verifying its ability to handle dynamic environments.