Monocular Depth Estimation Using HD Map Object Correlation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing monocular depth estimation algorithms suffer from scale ambiguity, limiting their ability to accurately determine metric distance measures from two-dimensional images, which is critical for applications like navigational assistance and autonomous vehicle control, especially when sensor failures occur.
Innovation Solution
Utilize high-definition map data to correlate objects within a two-dimensional image with precise map objects, determining pixel coordinates and camera parameters to estimate monocular depth scale, leveraging traditional image processing or AI-derived object detection, and incorporating inertial measurement units for relative pose changes.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If monocular depth estimation algorithms are used, then device complexity is reduced, but measurement precision of depth/distance deteriorates due to scale ambiguity
Solution Approach 1:
The patent introduces map data as an intermediary element that mediates between the monocular image and the depth estimation process. By correlating detected objects with their corresponding map objects, the system uses the known geometric information from HD maps to resolve the scale ambiguity inherent in monocular depth estimation, thereby improving measurement precision without adding complex sensors
Solution Approach 2:
The system changes the parameter basis for depth estimation by incorporating map-derived geometric parameters (such as lane widths, intersection dimensions, and object positions from HD maps) into the monocular depth estimation process. This parameter transformation allows the system to compute metric depth scales from single images by leveraging the known scale information embedded in map data
2Measurement precision
If traditional depth sensors (stereo cameras, LIDAR) are used, then measurement precision of depth is improved, but device complexity and cost increase
Solution Approach 1:
The patent creates a virtual depth map by copying and utilizing the geometric information already present in HD map data. Instead of relying on physical depth sensors, the system reconstructs depth information computationally by projecting map objects into the image space and using their known dimensions to infer metric scales, thereby achieving depth measurement precision without physical depth-sensing hardware
Solution Approach 2:
The system replaces mechanical/optical depth sensing mechanisms (stereo cameras, LIDAR) with a computational approach that uses monocular vision combined with HD map data. This substitution eliminates the need for complex sensor hardware while maintaining depth estimation capabilities through algorithmic processing of single images with map priors
3Measurement precision
If HD map data is correlated with image objects, then depth estimation accuracy is improved, but processing time increases
Solution Approach 1:
The system performs preliminary actions by pre-processing and storing geometric information in HD map data before runtime. Map objects, their positions, dimensions, and spatial relationships are pre-computed and structured in advance, allowing the real-time system to quickly correlate detected objects with pre-prepared map information rather than computing everything from scratch during operation
Data Source
AI summary
A method is provided for estimating depth within a two dimensional image. Methods may include: receiving a two dimensional image of an environment captured by an image sensor; extracting pixel coordinates of pixels associated with the one or more objects within the environment from the two dimensional image; correlating the one or more objects within the environment with the one or more map objects of map data; determining a depth of the pixels associated with the one or more objects within the environment based on a location of the one or more map objects in the map data relative to the image sensor; generating a depth map of the pixels associated with the one or more objects; and providing for at least one of navigational assistance or at least partial autonomous control to a vehicle based, at least in part, on the depth map.