Neural Network Localization Pose Training With Aerial Pre-Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Precise localization of mobile platforms in large geographical regions is challenging due to ambiguity in feature determination, especially when using high-resolution maps, and existing deep learning-based methods struggle with scalability and unambiguity.
Innovation Solution
A method is developed to train a neural convolutional network using both aerial and ground images, where aerial images are used for pre-training to learn spatial context, followed by ground image training cycles to minimize pose deviation, allowing for accurate localization without relying on high-resolution maps.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If high-resolution maps are used for localization, then measurement precision is improved, but device complexity and economic costs increase
Solution Approach 1:
The training process is segmented into two distinct phases: first training the neural network with aerial images to learn spatial context, then fine-tuning with ground images for precise pose determination. This segmentation allows the system to achieve high localization precision without requiring complex high-resolution maps during operation.
Solution Approach 2:
The neural network is pre-trained with aerial images before actual localization tasks. This preliminary training establishes spatial context and feature relationships that enable accurate localization without needing complex runtime map data, reducing device complexity while maintaining precision.
2Productivity
If deep learning-based methods are used for pose determination, then productivity is improved with constant query time, but measurement precision deteriorates in large geographical regions due to ambiguity
Solution Approach 1:
The solution transitions from two-dimensional ground images alone to incorporating aerial image perspectives during training. This dimensional change provides additional spatial context that disambiguates features in large geographical regions, improving pose determination accuracy while maintaining constant query time through the pre-trained network.
Solution Approach 2:
The neural network performs preliminary learning of spatial relationships from aerial images before actual pose determination. This preliminary action embeds contextual information that resolves ambiguities during runtime, enabling accurate localization across large regions without increasing query time.
3Device complexity
If ground images alone are used for training, then device complexity is reduced, but measurement precision worsens due to ambiguous feature allocation in large regions
Solution Approach 1:
The training data is segmented into two types: aerial images for learning spatial context and ground images for pose determination. This segmentation allows the system to achieve accurate feature allocation across large regions without requiring complex high-resolution maps during operation.
Solution Approach 2:
Aerial images serve as an intermediary training resource that provides spatial context without being used during actual localization. This intermediary training data helps the network learn to disambiguate ground image features, improving measurement precision while keeping runtime device complexity low.
Data Source
AI summary
A method for training a neural convolutional network for determining, with the aid of the neural convolutional network, a localization pose of a mobile platform using a ground image. Using a first multitude of aerial image training cycles, each aerial image training cycle includes: providing a reference pose of the mobile platform; and providing an aerial image of the environment of the mobile platform in the reference pose; using the aerial image as an input signal of the neural convolutional network; determining the respective localization pose with the aid of an output signal of the neural convolutional network; and adapting the neural convolutional network to minimize a deviation of the respective localization pose determined using the respective aerial image from the respective reference pose.