Pixel-Wise Motion Segmentation for SLAM in Dynamic Environments

Thorsten Hempel,Ayoub Al-Hamadi

doi:10.1109/access.2020.3022506

Abstract

Visual simultaneous localization and mapping (SLAM) is a key prerequisite for many mobile robotic systems. A common assumption for SLAM methods is a static environment. The interference of dynamic objects can lead to impairment of the camera pose tracking and permanent distortions of the map. This limits the use of many visual SLAM systems in real world scenarios, where dynamic environments are typical. We present a novel method for pixel-wise segmentation of dynamic image sequences based on a scene flow model estimation. We detect and eliminate outlying pixels sparsely by evaluating each pixel motion separately and maintain the most possible area of static scene background for SLAM. The evaluation with the public TUM dataset demonstrates that our proposed method outperforms other comparable state-of-the-art approaches for dynamic removal for SLAM systems.

Highlights

Simultaneous localization and mapping has been intensively researched over the past decades and became a fundamental capability for mobile robots
We show that our method of decoupling dynamic elements from the static background is remarkably efficient due its pixel-wise procedure and independence from further semantic analysis of the images (See Fig. 1)
In a third step we compare our results to other stateof-the-art slam systems for dynamic environments

Summary

Introduction

Simultaneous localization and mapping has been intensively researched over the past decades and became a fundamental capability for mobile robots. In visual SLAM the most simplistic setup is a monocular camera, which is popular due to its low cost, size and fast calibration. With the development of more complex sensors RGB-D and stereo cameras gained extensive popularity for enabling a metric scaling of the environment and led to improvements of the robustness of SLAM algorithms. A common reason is a flawed transformation estimation of the camera motion, as the matching between a frame with its previous reference is violated by dynamic objects. This significantly limits the use of many visual SLAM system in target applications, e.g. in mobile service robots and autonomous vehicles, where multiple dynamic elements in the environment are common

Methods

Results

Conclusion