CNN-SVO: Improving the Mapping in Semi-Direct Visual Odometry Using Single-Image Depth Prediction

Abstract

Reliable feature correspondence between frames is a critical step in visual odometry (VO) and visual simultaneous localization and mapping (V-SLAM) algorithms. In comparison with existing VO and V-SLAM algorithms, semi-direct visual odometry (SVO) has two main advantages that lead to stateof-the-art frame rate camera motion estimation: direct pixel correspondence and efficient implementation of probabilistic mapping method. This paper improves the SVO mapping by initializing the mean and the variance of the depth at a feature location according to the depth prediction from a singleimage depth prediction network. By significantly reducing the depth uncertainty of the initialized map point (i.e., small variance centred about the depth prediction), the benefits are twofold: reliable feature correspondence between views and fast convergence to the true depth in order to create new map points. We evaluate our method with two outdoor datasets: KITTI dataset and Oxford Robotcar dataset. The experimental results indicate that the improved SVO mapping results in increased robustness and camera tracking accuracy.

Simple But Effective Scale Estimation for Monocular Visual Odometry in Road Driving Scenarios

“…In addition, we also compare the performance of the proposed method with that of three CNN-based methods: Yin's method [16], Sucar's method [17], Yang's method [18], and Loo's method [19]. ORB-SLAM using the proposed scale estimation method with and without the guidance of the motion vector are referred to as GCSEMK (geometrically constrained scale estimation from multiple keyframes) and SEMK (scale estimation from multiple keyframes), respectively.…”

“…Comparing with the KITTI dataset, the RobotCar dataset is difficult for the vSLAM algorithms because of significant illumination changes. For some sequences in the Robot-Car dataset, the state-of-the-art vSLAM algorithms such as SVO [27], DSO [29], and ORB-SLAM cannot estimate camera pose due to feature tracking failure [19]. Since ORB-SLAM was used as our baseline algorithm, we tested the proposed method on the sequence for which the ORB-SLAM system can produce camera poses and 3D points.…”

“…To evaluate the performance of the scale drift elimination, we also computed the absolute trajectory error (ATE) [49] of the proposed method and compared it with those of ORB-SLAM, ORB-SLAM without loop closure (OS-NL), and ORB-SLAM with the ground-truth scale (OS-GTs). We followed the evaluation method used in Wang's [11] and Loo's method [19]. Since the absolute scale is unknown for the ORB-SLAM and OS-NL, for comparison, the ATE was computed after aligning the estimated trajectory with the ground truth by using a similarity transformation 3 .…”

“…Then, the resulting camera height is obtained by using the Kalman filter [8], [9] or pre-trained classifier [10]. Currently, the SLAM methods employing the information from deep learning such as the estimated depth and segmented region or 3D bounding box of an object have also been developed to lessen the scale drift [16]- [19].…”

DistanceNet: Estimating Traveled Distance From Monocular Images Using a Recurrent Convolutional Neural Network

“…Recent and state-of-the-art mono vSLAM/VO methods mainly use CNNs to resolve the scale ambiguity problem. The approaches of [19,26,25] and [14] have trained a CNN to deploy a scaled depth map from single monocular images. These dense depth maps are used to extend the optimization scheme of the classical frameworks, like for instance a so-called virtual stereo setup [25] for DSO or initializing depth filters [14] in SVO directly from the depth map, to eliminate the scale ambiguity.…”