WPO-Net: Windowed Pose Optimization Network for Monocular Visual Odometry Estimation

Abstract

Visual odometry is the process of estimating incremental localization of the camera in 3-dimensional space for autonomous driving. There have been new learning-based methods which do not require camera calibration and are robust to external noise. In this work, a new method that do not require camera calibration called the “windowed pose optimization network” is proposed to estimate the 6 degrees of freedom pose of a monocular camera. The architecture of the proposed network is based on supervised learning-based methods with feature encoder and pose regressor that takes multiple consecutive two grayscale image stacks at each step for training and enforces the composite pose constraints. The KITTI dataset is used to evaluate the performance of the proposed method. The proposed method yielded rotational error of 3.12 deg/100 m, and the training time is 41.32 ms, while inference time is 7.87 ms. Experiments demonstrate the competitive performance of the proposed method to other state-of-the-art related works which shows the novelty of the proposed technique.

From Pixels to Precision: A Survey of Monocular Visual Odometry in Digital Twin Applications

“…The issue of timestamp synchronization also emerges as a critical concern, as highlighted in [ 9 ]. Delays in timestamping due to factors like data transfer, sensor latency, and Operating System overhead can lead to discrepancies in visual–inertial measurements.…”

“…Timestamp Synchronization Uncertainty : This type of uncertainty arises when there are discrepancies in the timing of the data capture and processing among different components of a system, such as cameras, inertial measurement units (IMUs), and LiDAR scanners. In systems that rely on precise timing for data integration and analysis, such as visual–inertial navigation systems, this uncertainty can significantly impact accuracy [ 9 ].…”

“…Adding to the repertoire of semi-deep learning approaches, Ref. [ 9 ] introduced the Windowed Pose Optimization Network (WPO-Net) for VO estimation. In this method, features are extracted from input images, followed by relative pose computation, with a WPO-Net optimizing the pose over a sliding window.…”

“…Windowed optimization is a process that refines the local estimation of the camera trajectory by considering a certain number of previous frames or ‘window’ of frames. This approach helps to improve the accuracy of pose predictions derived from the analysis of image sequences [ 9 , 10 ].…”

An Unsupervised Monocular Visual Odometry Based on Multi-Scale Modeling

“…Compared to them, our method still achieves competitive performance. We also compare with the methods based on supervised learning [ 7 , 44 , 45 , 46 ]. Although these methods have simple network structure and fast training speed, they require ground truth to train the network.…”