Vehicle 3D Point Cloud Mapping with Lighting-Invariant Features
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Traditional map creation methods for autonomous driving, such as using panoramic surveillance systems, overhead views, and LiDAR point clouds, are not accurate and lack sufficient information for robust vehicle positioning due to sensitivity to lighting and view changes, and are inadequate in areas without landmarks.
Innovation Solution
A method using a neural network model trained on images under different lighting conditions to identify feature points and create 3D point cloud maps from image frames, enhancing robustness by incorporating stronger illumination and view invariance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional map creation methods (panoramic surveillance, overhead views, LiDAR) are used, then the system is simple to implement, but the positioning accuracy deteriorates under varying lighting and view conditions
Solution Approach 1:
The patent changes the parameter of feature representation from traditional geometric features to deep learning-based feature descriptors. By training neural networks on images under different lighting conditions, the system extracts robust features that maintain high positioning accuracy across varying environmental parameters without requiring complex additional hardware
Solution Approach 2:
The patent replaces traditional mechanical/optical mapping systems (panoramic cameras, overhead views, LiDAR) with a computational approach using deep learning models. The neural network-based feature extraction and matching system substitutes the physical sensor systems, achieving superior positioning accuracy through algorithmic processing rather than complex hardware configurations
2Reliability
If traditional feature matching methods are used, then the processing speed is fast, but the reliability deteriorates in areas without landmarks or under different lighting
Solution Approach 1:
The patent performs preliminary action by pre-training the neural network model on a large dataset of images under various lighting conditions before actual map creation. This pre-training ensures that when the system encounters new environments during positioning, the features are already optimized for robust matching, improving reliability without adding time during the actual positioning operation
Solution Approach 2:
The patent creates a universal feature extraction system that works across multiple conditions (different lighting, views, and environments) using a single trained neural network model. This multi-functional approach allows the same system to reliably match features in diverse scenarios without requiring condition-specific processing, maintaining both reliability and efficiency
3Measurement precision
If more detailed environmental information is collected to improve map accuracy, then the positioning precision improves, but the data processing complexity increases
Solution Approach 1:
The patent extracts only the essential and robust features from images using trained neural networks, rather than processing all environmental data. The feature extraction process selectively identifies key discriminative elements that are invariant to lighting and view changes, achieving high map accuracy while keeping data processing complexity manageable by filtering out redundant information
Data Source
AI summary
A method for creating a 3D point cloud map includes acquiring image frames from an image frame sequence of an external environment captured by a target camera of a vehicle. The method further includes processing the image frames using a feature point recognition model to identify a set of feature points in the image frames, along with a corresponding set of descriptors for the feature points. The feature point recognition model is a neural network model trained using a plurality of sample images of a same scene under different lighting intensities. The method further includes creating the 3D point cloud map of the external environment based on the set of descriptors. The method enables the obtained descriptors to include more image information, and further allows for an extraction of more matching feature points from the image when there are changes in lighting or view.