Robust ground plane detection for obstacle avoidance of mobile robots using a monocular camera

Abstract

This paper presents a vision-based obstacle avoidance design using a monocular camera onboard a mobile robot. An image processing procedure is developed to estimate distances between the robot and obstacles based-on inverse perspective transformation (IPT) in image plane. A robust image processing solution is proposed to detect and segment navigatable ground plane area within the camera view. The proposed method integrate robust feature matching with adaptive color segmentation for plane estimation and tracking to cope with variations in illumination and camera view. After IPT and ground region segmentation, a result similar to the occupancy grid map is obtained for mobile robot obstacle avoidance and navigation. Practical experimental results of a wheeled mobile robot show that the proposed imaging system successfully estimates distance of objects and avoid obstacles in an indoor environment.

Ground plane feature detection in mobile vision-aided inertial navigation

“…Regardless of the motion estimation uncertainties, which are nonlinearly propagated to the applied geometrical constraints, the dynamics of the mobile vision system may cause conditions for which the preceding constraints are not sufficient for correct detection; that is, image features that do not lie in the same physical plane share the same planar homography which in turn will result in an incorrect detection, i.e., the virtual plane false detection [9], [11]. In short, the virtual plane false detection happens when the camera relative translation t is close to zero.…”

“…Moreover, the normalbased outlier rejection is not able to obtain a valid estimate of the ground plane normal vector because the matrix S in (15) approaches the zero matrix. Based on the estimated motion parameters, we perform a simple stationarity test by examining the camera displacement among the consecutive frames, similar to [9]. Hence whenever a stationary phase is encountered (t ≈ 0), the previous image is replaced with a preceding image for which the camera relative displacement is not close to zero.…”

“…Vision-based ground plane detection frameworks are generally based on a single camera that can be used together with complementary sensors, such as a secondary camera (stereo-vision) [1]- [3], laser scanners [3], or inertial sensors [1], to increase the flexibility or performance of the system. Depending on the hardware structure, vision-based ground plane detection approaches benefit from a priori knowledge of the scene or system structure; for instance, the geometric assumptions on the nature of the traffic environments [3]- [5], planar properties of indoor environments [6], color cue segmentations [7]- [9], and the assumption of the ground plane projected region on the image plane [4]. For instance, in [4] a method for road detection and obstacle detection is presented that is entirely based on stereovision.…”

“…Virtual plane false detection happens when a set of feature points that share the same homography are not lying on the same physical plane. Only few studies have addressed the effects of virtual plane false detection or how to overcome this issue [9], [11]. For instance, Zhou and Li [11] have introduced an ad-hoc method based on the color classification to avoid the virtual plane feature detection.…”

“…Additionally, a second-step outlier rejection method is proposed based on a new ground-plane normal computation approach using only three pairs of corresponding feature points. Furthermore, similar to [9], a simple stationarity test is performed to avoid virtual plane false detection. This additional step significantly improves the performance of the detection algorithm when the camera is stationary.…”