Moving object detection for camera pose estimation in dynamic environments

The reported simultaneous localization and mapping (SLAM) methods show satisfactory object detection results in static environments. However, most of them cannot be directly applied in the moving object detection in various environments due to high detection errors. This work is focused on developing a new method to solve the problem of moving object detection under dynamic environment. To this end, a three-stage program is proposed based on the RGB-D image acquisition. Firstly, target objects are detected by the YOLOv3 model from the RGB images. Secondly, the object features are clustered using the k-means algorithm by fusing the information of the depth images. Finally, the moving objects are identified under dynamic environment based on the multi-view geometry theory. Experimental results show that the localization accuracy of SLAM in dynamic environment can be improved obviously by using the moving object detection method in this paper.

I. Introduction

SLAM was proposed in 1986 [1], and has been an important technology for robotic applications in the decades [2]–[4]. For example, ORB-SLAM2 is a developed SLAM model by integrating ORB feature extraction algorithm to support RGB-D camera applications [5], [6]. However, the ORB-SLAM2 is focused on the map-building in static environment. The camera pose estimation and the environment mapping accuracy are affected by moving objects. The mapping errors also will increase during the period of SLAM implemented, leading to incorrect object detection and location.