Binocular Camera Map Update for Autonomous Driving
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Updating high-resolution maps for autonomous driving systems is economically and practically challenging due to the high cost of dispatching LiDAR-equipped survey vehicles for frequent infrastructure changes, and low-resolution data from monocular cameras lacks depth information necessary for accurate 3D modeling.
Innovation Solution
A system utilizing binocular cameras mounted on vehicles to capture depth information and pose data, which is processed to generate point cloud frames, merge them with existing high-resolution maps, and determine if the updated area requires further LiDAR survey based on size thresholds.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If LiDAR-equipped survey vehicles are dispatched to update high-resolution maps, then map accuracy and depth information are improved, but operational cost increases significantly
Solution Approach 1:
The patent segments the map update process into two stages: initial high-resolution mapping using LiDAR, and subsequent incremental updates using inexpensive monocular cameras. This segmentation allows the expensive LiDAR system to be used only when necessary (initial mapping), while routine updates utilize low-cost cameras, thereby reducing operational costs while maintaining map accuracy
Solution Approach 2:
The patent performs preliminary high-resolution mapping using LiDAR to create an initial accurate map. This preliminary action establishes a baseline that enables subsequent use of cheaper monocular cameras for incremental updates, as the camera-based system can leverage the pre-established 3D geometry and depth information from the initial LiDAR survey
2Productivity
If monocular cameras are used to update maps, then operational cost is reduced, but depth information and 3D modeling quality deteriorate
Solution Approach 1:
The patent introduces SfM (Structure from Motion) algorithms as an intermediary processing step that transforms 2D monocular camera images into 3D point clouds. This intermediary computational process enables the extraction of depth information and 3D geometry from inexpensive monocular camera data, allowing cost-effective updates while maintaining adequate depth information quality
Solution Approach 2:
The patent creates a virtual copy of the 3D environment using SfM algorithms that processes monocular images to generate point cloud representations. This computational copying approach reconstructs depth information and 3D structure from 2D images, enabling cost-effective map updates without direct LiDAR measurement while preserving essential spatial information
3Reliability
If frequent map updates are performed to reflect infrastructure changes, then map relevance is improved, but the frequency and cost of LiDAR surveys increase
Solution Approach 1:
The patent segments the update frequency into two tiers: high-frequency incremental updates using monocular cameras for routine infrastructure changes, and low-frequency LiDAR surveys for comprehensive re-mapping. This segmentation enables frequent map updates to maintain relevance while avoiding the high cost of frequent LiDAR deployments
Solution Approach 2:
The patent implements a dynamic update strategy where the system automatically selects between monocular camera-based updates and LiDAR-based updates based on the nature and scale of detected changes. This dynamic approach allows the system to adaptively choose the appropriate update method, maintaining map relevance while optimizing survey frequency and cost
Data Source
Figure 1
Figure 2
Figure 3
AI summary
Embodiments of the disclosure provide systems and methods for updating a high-resolution map. The system may include a communication interface configured to receive a plurality of image frames captured by a binocular camera equipped on a vehicle, as the vehicle travels along a trajectory. The system may further include a storage configured to store the high-resolution map and the plurality of image frames. The system may also include at least one processor. The at least one processor may be configured to generate point cloud frames based on the respective image frames. The at least one processor may be further configured to position the vehicle using the point cloud frames. The at least one processor may be further configured to merge the point cloud frames based on the vehicle positions. The at least one processor may also be configured to update a portion of the high-resolution map based on the merged point cloud.