Robust Localization System Fusing Vision and Lidar Under Severe Occlusion

Yongliang Shi,Weimin Zhang,Fangxing Li,Qiang Huang

doi:10.1109/access.2020.2981520

Abstract

Localization is one of the most fundamental problems for mobile robot. Aiming at the phenomenon that robot is prone to be lost in navigation under severe occlusion, a robust localization system combining vision and lidar is proposed in this paper. The system is split into off-line stage and online stage. In the off-line stage, this paper introduces a method of actively detecting and recording visual landmarks, and an off-line visual bag-of-words is generated from the recorded landmarks training. In the online stage, the prediction and update phase of Adaptive Monte Carlo Localization (AMCL) are improved respectively to enhance the performance of localization. The prediction phase generates the proposal distribution according to the prior information obtained through retrieving visual landmarks, and the newly proposed measurement model that selects reliable beams of lidar as the observation is to update the prediction. Experiments is carried out under strict conditions, that is 60% of the lidar is occluded, 1/12 of the beams are regarded as observation, and only 300 particles were adopted at most, it is shown that, no matter in the global localization or pose tracking, the localization system proposed in this paper performs much better than the state of art localization algorithm AMCL.

Highlights

Localization against a global map, including global localization and pose tracking, is a prerequisite for any robotics task where a robot must know where it is
The Markov localization method requires excessive computational overhead and a priority commitment to the size and resolution of the state space [6], which is difficult to meet the real-time requirements; The localization efficiency based on Kalman filter (KF) is high, but the Kalman filter is based on the assumption of linear Gaussian model, and it is difficult to describe the motion model and observation model accurately with linear model; On this basis, extended Kalman filter (EKF) and unscented Kalman filter algorithm are introduced [7], [8], these two methods solve the state transfer by linearizing the non-linear model of the original system near the operating point by first-order Jacobian approximation [9]
Experiments are conducted for global localization and pose tracking under severe occlusion

Summary

Introduction

Localization against a global map, including global localization and pose tracking, is a prerequisite for any robotics task where a robot must know where it is. The Markov localization method requires excessive computational overhead and a priority commitment to the size and resolution of the state space [6], which is difficult to meet the real-time requirements; The localization efficiency based on Kalman filter (KF) is high, but the Kalman filter is based on the assumption of linear Gaussian model, and it is difficult to describe the motion model and observation model accurately with linear model; On this basis, extended Kalman filter (EKF) and unscented Kalman filter algorithm are introduced [7], [8], these two methods solve the state transfer by linearizing the non-linear model of the original system near the operating point by first-order Jacobian approximation [9] These two methods perform well in the pose estimation of the robot under certain. The it is easy to introduce erroneous observations under severe occlusion conditions leading to failure of localization

Objectives

Results

Conclusion