Autonomous Vehicle Localization via Visual Place Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Autonomous vehicles face challenges in precise localization due to the limitations of commercial GPS solutions, which have low accuracy and are prone to interference in urban areas with high-rise buildings, necessitating alternative methods for self-localization within a few centimeters.
Innovation Solution
The integration of visual place recognition into digital maps using camera sensors and particle-filtering techniques, where feature vectors are continuously updated and matched in real-time to aid in precise vehicle localization, combining with visual odometry for efficient self-localization.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If GPS is used for vehicle localization, then the system is simple and provides coverage everywhere, but the accuracy is only a few meters and signals are interfered with by large buildings, canyons, trees, and power lines
Solution Approach 1:
The patent combines multiple localization approaches (GPS, visual odometry, and visual place recognition) into a unified system. The visual place recognition system merges feature extraction from camera images with particle filtering and digital map matching to achieve centimeter-level accuracy, while GPS provides coarse positioning. This combination resolves the contradiction by integrating simple and complex methods to achieve both accuracy and robustness.
Solution Approach 2:
The patent introduces visual features and digital maps as intermediary elements between the vehicle and the localization goal. Instead of relying directly on GPS signals, the system uses camera-captured images, extracts visual features, compares them with pre-stored digital map features, and determines position through feature matching. This intermediary approach enables precise localization independent of GPS signal quality.
2Measurement precision
If visual place recognition with feature vector matching is implemented, then localization accuracy improves to within a few centimeters, but computational complexity and processing requirements increase
Solution Approach 1:
The patent performs preliminary actions by pre-processing and storing visual features in digital maps before the vehicle reaches those locations. During localization, the system only needs to compare current camera features with pre-stored features rather than processing raw images from scratch. This advance preparation significantly reduces real-time computational complexity while maintaining high localization accuracy.
Solution Approach 2:
The patent extracts only the essential visual features from images rather than processing complete image data. By identifying and extracting key visual landmarks and characteristics, the system reduces the data volume that needs to be compared and processed, thereby lowering computational complexity while preserving localization precision.
3Measurement precision
If real-time feature vector updating and matching is performed, then localization precision reaches within a few centimeters, but processing time and computational resources increase
Solution Approach 1:
The patent applies partial action by updating and matching only the most relevant feature vectors rather than processing all possible features. The particle filtering mechanism focuses computational efforts on the most probable location hypotheses, comparing features only in those regions. This selective approach achieves centimeter-level precision while minimizing processing time by avoiding exhaustive computation.
4Measurement precision
If GPS reliance is reduced for autonomous driving precision, then localization accuracy within lanes improves, but the system becomes more complex requiring camera sensors and visual processing
Solution Approach 1:
The patent makes the camera sensor multi-functional by using it for both visual odometry (motion estimation) and visual place recognition (position determination). Instead of adding separate dedicated sensors for each function, the same camera system serves multiple purposes: capturing images for feature extraction, tracking visual motion, and comparing with digital maps. This universality reduces overall system complexity while achieving the required position accuracy for lane-level autonomous driving.
Data Source
AI summary
Methods and systems herein can let an autonomous vehicle localize itself precisely and in near real-time in a digital map using visual place recognition. Commercial GPS solutions used in the production of autonomous vehicles generally have very low accuracy. For autonomous driving, the vehicle may need to be able to localize in the map very precisely, for example, within a few centimeters. The method and systems herein incorporate visual place recognition into the digital map and localization process. The roadways or routes within the map can be characterized as a set of nodes, which can be augmented with feature vectors that represent the visual scenes captured using camera sensors. These feature vectors can be constantly updated on the map server and then provided to the vehicles driving the roadways. This process can help create and maintain a diverse set of features for visual place recognition.