Multi-Camera BEV Path Planning Without GPS or LiDAR
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing path-planning methods for autonomous vehicles rely on external localization devices like GPS or LiDAR, which are costly, require specialized hardware, and struggle with accuracy and scalability, especially in racing scenarios where precise localization is needed.
Innovation Solution
A system using onboard cameras and neural networks to generate a Bird's Eye View (BEV) costmap, which is used by a path-planning module to calculate optimal trajectories without external localization, leveraging convolutional and transformer-based neural networks for feature extraction and spatial correlation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If GPS is used for positioning, then cost is reduced, but positioning accuracy deteriorates to +/- 1 to 3 meters
Solution Approach 1:
The patent introduces visual features from the environment as an intermediary to bridge the gap between low-cost GPS and high-precision localization. By extracting and matching visual features (edges, corners, textures) from camera images with pre-built maps, the system achieves centimeter-level positioning accuracy without requiring expensive GPS hardware, thus resolving the contradiction between cost and accuracy
Solution Approach 2:
The patent creates a visual copy of the environment by capturing images and extracting visual features to build a map. This visual map serves as a replica of the physical environment that can be matched against real-time camera inputs, enabling accurate localization through image correlation rather than relying on expensive positioning hardware
2Measurement precision
If LiDAR is used for localization, then positioning accuracy is improved, but cost and system complexity increase
Solution Approach 1:
The patent replaces the mechanical LiDAR scanning system with an optical-based vision system. Instead of using laser ranges to measure distances, the system uses camera images and visual feature extraction to achieve localization. This substitution eliminates the need for complex LiDAR hardware while maintaining positioning accuracy through computational methods
Solution Approach 2:
The patent creates a visual representation copy of the environment by extracting visual features from camera images and storing them in a pre-built map. This visual copy allows the system to perform localization by matching current image features with stored map features, replacing the need for LiDAR point cloud processing and simplifying the overall system
3Productivity
If pre-generated costmaps are used, then runtime computation is reduced, but localization precision requirements increase
Solution Approach 1:
The patent performs preliminary action by pre-generating costmaps and extracting visual features from training images during an offline phase. This pre-processing creates a ready-to-use map structure with embedded visual features, allowing the runtime system to quickly match current images with pre-computed costmaps without requiring high-precision localization algorithms, thus resolving the contradiction between computation efficiency and precision requirements
Data Source
AI summary
Systems and methods for autonomous-vehicle navigation integrating path planning with a perception network. A Bird's Eye View costmap is generated at runtime using only onboard sensors. No external localization providers are used.