Vanishing point detection in corridor for autonomous mobile robots using monocular low-resolution fisheye vision

Dejing Ni,Aiguo Song,Peng Ji

doi:10.1177/1687814019884767

Dejing Ni, Aiguo Song + Show 1 more

Open Access

https://doi.org/10.1177/1687814019884767

Copy DOI

Abstract

It is crucial for mobile robots to implement vanishing point detection during navigation in corridors. For the fisheye vision, the conventional methods of vanishing point detection usually obtain poor detection results. This is mainly attributed to serious barrel distortion in images acquired from fisheye cameras that are widely used in mobile robot systems. In the proposed system, a novel vanishing point detection algorithm based on the Gabor filter bank and the convolutional neural network is put forward to realize more accurate detection. The Gabor filter bank is used to extract image texture information in the preprocessing step, thereby enhancing the generalization. The convolutional neural network is used to predict the position of the vanishing point in the fisheye images. To improve the real-time performance and guarantee the accuracy, the low-resolution image should be selected as the input image as far as possible. For this purpose, a multi-resolution experiment was carried out. With the appropriate resolution, the proposed vanishing point detector was found still effective even if 60% of the original information was discarded. In addition, an experiment was conducted to verify the generalization on the condition of illumination changing, pedestrians passing, and different corridor appearance. The experiments displayed good effect and generalization on fisheye images captured in the corridor.

Highlights

Autonomous navigation in the corridor using monocular vision has aroused a wide range of attention in the field of mobile robots
In a projected image plane, those parallel lines may no longer parallel but converge to one point that is defined as the vanishing point (VP)
Providing important cues for inferring the three-dimensional (3D) structure of a real scene, VP is widely applied in mobile robot navigation,[1] road detection[2,3] and camera calibration,[4] and so on

Summary

Introduction

Autonomous navigation in the corridor using monocular vision has aroused a wide range of attention in the field of mobile robots. Taking the monocular camera as the navigation camera for a mobile robot, this article proposes a novel VP detection algorithm based on the supervised method, which requires only a sample set marked with the location of VP. Different from the traditional edge-based methods, texture-based methods, and optimization-based methods, our VP detection method takes the VP detection as a regression problem For this purpose, a novel VP detector based on a convolutional neural network is proposed, which consists of two parts, the image texture information extraction and the merging of the predictions. To eliminate the interference of vertical segments in the image and enhance generalization ability of the VP detector, a Gabor filter bank (five scales at seven directions, not including 90 degree) is employed to extract the texture feature information first. The final VP location is represented by the mean value of the five prediction results, as shown in the following

Experiments

Findings

Conclusion