Vehicle Localization Using Layered Hypothesis Filters
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional vehicle localization methods in complex road networks, such as those with elevated roads, are often inaccurate and computationally expensive due to the lack of height information and regional restrictions on HD maps and GNSS data, leading to challenges in determining the vehicle's position and orientation.
Innovation Solution
A computer-implemented method that calculates a combined probability value for candidate states based on previous states, transition probabilities, emission probabilities, and elevated-road ramp probabilities to determine the most likely sequence of states, allowing for accurate and efficient localization of vehicles in complex road networks.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional localization algorithms are used with HD maps and sensor fusion, then vehicle localization can be achieved, but the system becomes computationally expensive and inaccurate in complex road networks with elevated roads
Solution Approach 1:
The road network is segmented into multiple layers (ground level and elevated levels) with distinct representations. Each layer is processed independently through separate hypothesis filters, allowing the system to handle complex 3D road networks while maintaining computational efficiency by avoiding the need to process all possible road combinations simultaneously.
Solution Approach 2:
The patent introduces a vertical dimension to the traditional 2D map matching approach by representing elevated roads as separate layers in the hypothesis filter. This dimensional separation allows the system to disambiguate between ground-level and elevated-road intersections, improving localization accuracy without proportionally increasing computational burden.
2Reliability
If multiple hypothesis filters are formed based on rough initial position estimate, then localization can be performed, but the number of hypotheses increases significantly in complex scenarios leading to higher computational cost
Solution Approach 1:
The system dynamically adjusts the number and characteristics of hypothesis filters based on the local road network structure and vehicle state. In complex areas with elevated roads, additional layer-specific filters are created, while in simpler areas fewer filters are needed. This local adaptation maintains robustness where needed while reducing overall computational complexity.
Solution Approach 2:
The hypothesis filter system is made dynamic by continuously updating and pruning filters based on sensor observations and vehicle motion models. Filters that contradict observed data are eliminated, and the system adapts the number of active filters based on the current localization uncertainty and road network complexity, preventing fixed high computational costs.
3Adaptability or versatility
If height information is unavailable or restricted, then regional restrictions are satisfied, but the ability to distinguish between overlaying roads is lost leading to inaccurate localization
Solution Approach 1:
The patent introduces an intermediary representation layer that models the 3D spatial relationships between ground-level and elevated roads without requiring direct height measurements from sensors. This intermediary graph structure allows the system to infer vertical positions and distinguish between overlaying roads using only 2D map data and vehicle odometry, maintaining regional compliance while achieving accurate localization.
Data Source
AI summary
A method for localizing a vehicle on a road is disclosed. The method includes for a time step out of a plurality of consecutive time steps, obtaining a set of candidate states for the vehicle on the road. Then for each candidate state, determining a probability of the vehicle being in that candidate state based on a combined probability value. The method includes determining a sequence of candidate states, over the plurality of consecutive time steps, which is associated with a highest probability out of a plurality of possible sequences of candidate states, wherein each of them includes one candidate state from each time step of the plurality of consecutive time steps. The method further includes outputting the road that the vehicle currently is on or a lane that the vehicle currently is in based on the determined sequence of states that is associated with the highest probability.