Topological Map Navigation for GPS-Constrained Autonomous Driving
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing autonomous driving techniques rely on high-definition maps, which are time-consuming and labor-intensive to acquire and maintain, and map-free methods face challenges in navigating complex environments due to limited action decisions, reliance on images affected by light conditions, and difficulty in training robust navigation models.
Innovation Solution
A smart navigation method and system based on a topological map that determines a travelable region view using a multi-index navigation model, incorporating scene data from cameras and lidar, and employs deep reinforcement learning to predict action decisions, including travel speed, direction, and deflection angles, under GPS constraints.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If high-definition maps are used for smart navigation, then navigation accuracy is improved, but map acquisition and maintenance become time-consuming and labor-intensive
Solution Approach 1:
The patent extracts only the essential topological information (road connections, intersections, travelable regions) from complete high-definition maps, creating a simplified topological map that retains navigation functionality while eliminating the need for comprehensive map acquisition and maintenance
Solution Approach 2:
The patent uses lightweight, easily updatable topological map data that can be quickly generated and modified without the substantial time and resource investment required for traditional high-definition maps, effectively treating the map data as a disposable resource that can be regenerated as needed
2Loss of time
If map-free navigation by street scenes is used, then map acquisition time is reduced, but action decision capability is limited and cannot cope with real driving
Solution Approach 1:
The patent performs preliminary construction of topological maps and pre-computation of travelable regions before actual navigation tasks, enabling the system to make more sophisticated action decisions in real-time without the overhead of complex map processing during driving
Solution Approach 2:
The patent transitions from the limited 5-action space of prior map-free methods to a continuous action space by integrating topological map reasoning with deep reinforcement learning, allowing the vehicle to select from infinitely many combinations of speed, direction, and deflection angle based on the current state and target location
3Device complexity
If complete reliance on images is used for navigation, then sensor requirements are simplified, but navigation performance deteriorates in weak light conditions
Solution Approach 1:
The patent creates a multi-functional navigation system where the topological map provides location and route guidance independent of lighting conditions, while scene images provide visual context, with the two sources working together to ensure reliable navigation in all environmental conditions
4Device complexity
If no map is used for navigation, then system simplicity is improved, but state space expansion makes optimization difficult and training time-consuming
Solution Approach 1:
The patent segments the navigation problem into two parts: topological map-based route planning (which reduces the state space by providing structured road network information) and local navigation decision-making (handled by deep reinforcement learning), making the overall optimization process more efficient and trainable
Data Source
AI summary
The invention discloses a smart navigation method and system based on a topological map, and relates to the technical field of computers. The smart navigation method based on a topological map comprises: determining a travelable region view according to current location information based on a constructed topological map (S101); acquiring scene data, the scene data at least including a scene image, a scene depth map, and a scene analysis map (S102); and determining an action decision based on the travelable region view, the scene data, and a navigation model (S103). The travelable region view is determined based on a multi-index navigation model and the constructed topological map; and relative to a map-free mode, the accuracy of real-time navigation can be improved under constraint of GPS, but GPS is not entirely relied on, so robustness of navigation can be improved.