Deep Odometry Neural Network for Visual Motion Estimation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing autonomous vehicle datasets are not well generalized across different environments, and feature-based methods fail when scenes lack salient keypoints, making it challenging to process and analyze data efficiently for visual odometry in autonomous vehicles.
Innovation Solution
A method and system for deep odometry using a neural network that performs data alignment among LiDAR, cameras, and IMU-GPS modules, generating point clouds, and predicting optical flow using convolutional neural networks (CNNs), deconvolution neural networks (DNNs), and recurrent neural networks (RNNs) to estimate motion parameters without relying on hand-crafted features.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If hand-crafted features are employed to extract keypoints and descriptors, then motion parameters can be solved through matching points, but the method fails when a scene has no salient keypoints
Solution Approach 1:
The patent replaces hand-crafted feature extraction methods with a neural network-based deep learning system. The neural network automatically learns feature representations from image data, eliminating the need for manual keypoint detection and descriptor matching. This substitution enables the system to handle diverse scene types including those without salient keypoints, as the network adapts its feature extraction based on learned patterns rather than relying on pre-defined mechanical feature detection algorithms
Solution Approach 2:
The patent transforms the approach by changing from fixed hand-crafted feature parameters to dynamic learned parameters through neural network training. The system learns optimal feature representations and motion parameter estimations from training data, allowing it to adapt to different environments and scene types. This parameter transformation enables the system to maintain reliability across varying conditions by learning context-appropriate features rather than relying on universal hand-crafted descriptors
2Ease of manufacture
If existing autonomous vehicle datasets are used, then data collection is feasible, but the datasets are not well generalized to different environments
Solution Approach 1:
The patent applies preliminary action through extensive training of the neural network on diverse datasets before deployment. The system is pre-trained on multiple autonomous vehicle datasets (such as KITTI, Waymo, and nuScenes) to learn robust feature representations and motion estimation patterns. This preliminary training enables the model to generalize to different environments without requiring re-collection of data for each new scenario, thus maintaining ease of deployment while improving adaptability
Solution Approach 2:
The patent creates a universal neural network model that can handle multiple functions: feature extraction, optical flow estimation, and motion parameter calculation. This multi-functional system is trained on diverse datasets to learn environment-agnostic representations, enabling it to generalize across different scenes, weather conditions, and lighting scenarios. The universal model eliminates the need for environment-specific feature extraction methods, achieving both ease of data collection and broad generalization
3Adaptability or versatility
If neural networks are used for deep odometry, then generalization ability is enhanced, but the device complexity increases
Solution Approach 1:
The patent segments the neural network architecture into distinct functional modules: a convolutional neural network (CNN) for feature extraction from image data, a deconvolutional neural network (DCNN) for optical flow estimation, and a recurrent neural network (RNN) for temporal sequence processing and motion parameter prediction. This segmentation allows each module to be optimized independently and trained separately, reducing the overall complexity while maintaining generalization ability. The modular structure also enables efficient computation by processing data in discrete stages rather than requiring a monolithic complex network
Data Source
AI summary
A method of visual odometry for a non-transitory computer readable storage medium storing one or more programs is disclosed. The one or more programs includes instructions, which when executed by a computing device, causes the computing device to perform the following steps comprising: performing data alignment among sensors including a LiDAR, cameras and an IMU-GPS module; collecting image data and generating point clouds; processing, in the IMU-GPS module, a pair of consecutive images in the image data to recognize pixels corresponding to a same point in the point clouds; and establishing an optical flow for visual odometry.