Crowd-Sourced HD Map Updating With Sparse Context-Aware Stitching
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current navigation systems for autonomous vehicles face challenges in maintaining precise and up-to-date map details, especially in dynamic environments, due to limitations in wireless communication protocols and the need for frequent updates incorporating data from various sources.
Innovation Solution
The system employs an on-vehicle navigation map updated through a spatial monitoring system that captures 3D sensor representations, executes feature extraction and semantic segmentation, and uses SLAM to integrate context, with a parsimonious map representation communicated to an off-board controller for cloud-based stitching and updating of a base navigation map.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If frequent map updates are performed to maintain precision and up-to-date details, then navigation map accuracy is improved, but wireless communication bandwidth and system resources are overwhelmed
Solution Approach 1:
The system extracts only the essential and changed features from the environment rather than transmitting complete map data. The off-board controller executes feature extraction routines that identify and isolate relevant environmental features (road markings, signs, obstacles) from sensor data, transmitting only these extracted features for map updates, thereby reducing data volume while maintaining accuracy
Solution Approach 2:
The navigation map is segmented into multiple components including static map data, dynamic features, and contextual information. The system updates only the specific segments that have changed rather than refreshing the entire map, allowing frequent updates without overwhelming communication bandwidth
2Measurement precision
If complete environmental details are captured and transmitted for map updates, then map detail precision is improved, but communication efficiency and bandwidth utilization deteriorate
Solution Approach 1:
The system employs feature extraction routines that isolate and identify only the relevant environmental features needed for navigation (road markings, traffic signs, obstacles, intersections) from the complete sensor data stream. This selective extraction transmits only essential information rather than complete environmental details, maintaining map precision while improving communication efficiency
Solution Approach 2:
The system transmits a partial representation of the environment focused on navigation-critical features rather than complete environmental data. The feature extraction process selectively captures and transmits only the portion of environmental information necessary for safe and accurate navigation, achieving sufficient precision without excessive data transmission
3Measurement precision
If real-time map updates are performed using all available sensor data, then navigation precision is improved, but computational load and processing time increase
Solution Approach 1:
The computational process is segmented into distinct stages: feature extraction, semantic segmentation, SLAM processing, and map updating. Each stage processes only the relevant data for its specific function, distributing computational load across multiple specialized routines rather than processing all sensor data through a single complex pipeline
Solution Approach 2:
The system performs preliminary feature extraction and semantic segmentation on sensor data before executing SLAM and map updating routines. By pre-processing and organizing sensor data into extracted features with semantic labels, the system reduces the computational complexity of subsequent navigation-critical operations
Data Source
AI summary
A vehicle is described, and includes an on-board controller, an extra-vehicle communication system, a GPS sensor, a spatial monitoring system, and a navigation system that employs an on-vehicle navigation map. Operation includes capturing a 3D sensor representation of a field of view and an associated GPS location, executing a feature extraction routine, executing a semantic segmentation of the extracted features, executing a simultaneous location and mapping (SLAM) of the extracted features, executing a context extraction from the simultaneous location and mapping of the extracted features, and updating the on-vehicle navigation map based thereon. A parsimonious map representation is generated based upon the updated on-vehicle navigation map, and is communicated to a second, off-board controller. The second controller executes a sparse map stitching to update a base navigation map based upon the parsimonious map representation. The on-vehicle navigation map is updated based upon the off-board navigation map.