Robust edge-based visual odometry using machine-learned edges

Abstract

In this work, we present a real-time robust edgebased visual odometry framework for RGBD sensors (REVO). Even though our method is independent of the edge detection algorithm, we show that the use of state-of-the-art machinelearned edges gives significant improvements in terms of robustness and accuracy compared to standard edge detection methods. In contrast to approaches that heavily rely on the photo-consistency assumption, edges are less influenced by lighting changes and the sparse edge representation offers a larger convergence basin while the pose estimates are also very fast to compute. Further, we introduce a measure for tracking quality, which we use to determine when to insert a new key frame. We show the feasibility of our system on realworld datasets and extensively evaluate on standard benchmark sequences to demonstrate the performance in a wide variety of scenes and camera motions. Our framework runs in real-time on the CPU of a laptop computer and is available online.

Combination of Planes and Image Edges for a Visual Odometer for Weakly Textured Structured Environments

“…The information about the environment is introduced in our system as camera frames, one colour image and one depth image. Next, each frame is processed by the Robust Edge-Based Visual Odometry (REVO) algorithm [15], which extracts edges, detects keyframes and estimates the transformation between keyframes. When a new keyframe is generated, we extract a coloured point cloud by combining the information from the colour and the depth images, detect [16] and describe the existing planes, and match them using our plane matcher.…”

Combining Edge Images and Depth Maps for Robust Visual Odometry

“…In contrast to [7,14], LSD-SLAM [4] is a semi-dense method that only relies on high-gradient regions to minimize the photoconsistency error. However, the quality of photoconsistency-based pose estimation suffers under motion blur and as demonstrated in [8,13] is limited to small inter-frame motions. Instead of high-gradient regions, many methods use the more robust edges.…”

“…Wang et al [17] jointly minimize edge distance and a photometric error at high-gradient pixels. In [13], Schenk and Fraundorfer study the influence of different edge detectors and demonstrate how to efficiently remove outliers to increase accuracy and robustness. Similar to feature-based methods, the distribution of edges can influence pose estimation.…”

“…In our optimization formulation, cost evaluation solely depends on the reference frame, thus we have to recompute all frame-dependent structures only when a new reference frame is added. We follow the strategy proposed in [13] and insert a new reference frame, when the overlap between reprojected edges from 3 previous frames and the edges in the current frame is lower than the number of non-overlapping edges.…”

“…In contrast, (ii) direct methods work better in poorly textured environments as they do not rely on feature extraction and matching but process image information directly. Dense or semi-dense approaches based on the photoconsistency assumption [4,7,14] and edge-based variants [8,13] are used. (iii) Iterative closest point (ICP) methods [10,18] directly align 3D point clouds but require sufficient 3D structure and a costly correspondence matching step.…”