Unstructured Lane Estimation Using Sensor Traces and Overhead Maps
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Estimating drivable areas within unstructured environments, such as parking lots, is difficult due to the lack of explicit lane structures in traditional maps.
Innovation Solution
A system utilizing a processor and memory to receive overhead representations and rasterized traces from vehicle sensors, applying a machine learning model to generate probabilities that neighboring pixels are part of a lane, which are then used to create lane graphs.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If traditional maps are used for lane estimation, then mapping accuracy is maintained in structured areas, but lane estimation becomes impossible in unstructured areas without explicit lane markings
Solution Approach 1:
The patent transitions from 2D overhead map representations to 3D point cloud data from LiDAR sensors, adding vertical dimension information. This enables the system to detect drivable areas in unstructured environments by analyzing spatial point distributions rather than relying on flat map images, thus achieving lane estimation capability where traditional maps fail.
Solution Approach 2:
The patent introduces an intermediary processing layer that converts raw LiDAR point cloud data into structured drivable area representations. This intermediary step bridges the gap between unstructured sensor data and the structured output needed for lane estimation, enabling reliable mapping in unstructured areas by mediating between the two data types.
2Manufacturing precision
If manual methods are used to create HD maps, then mapping accuracy is high, but productivity is significantly reduced due to laborious processes
Solution Approach 1:
The patent implements a self-service mapping system where vehicles automatically collect, process, and generate map data during normal operation. The system uses onboard LiDAR sensors to autonomously capture point cloud data, processes it through machine learning models to identify drivable areas, and updates maps without human intervention, thereby maintaining high accuracy while dramatically improving productivity.
Solution Approach 2:
The patent replaces manual mechanical map creation processes with automated computational systems. Instead of workers manually drawing and annotating maps, the system uses LiDAR technology and machine learning algorithms to automatically generate accurate drivable area representations, substituting human labor with automated sensing and processing.
3Measurement precision
If LiDAR point cloud data is processed using traditional methods, then computational simplicity is maintained, but measurement precision of drivable areas deteriorates in unstructured environments
Solution Approach 1:
The patent introduces intermediary processing steps that convert raw LiDAR point cloud data into structured representations suitable for analysis. This includes creating overhead views, generating depth maps, and extracting geometric features as intermediate products that bridge the gap between raw sensor data and final drivable area detection, improving precision without overwhelming computational complexity.
Solution Approach 2:
The patent segments the LiDAR point cloud data into meaningful components such as ground points, vegetation, buildings, and potential drivable areas. By dividing the complex point cloud into manageable segments and processing each separately, the system achieves high measurement precision for drivable areas while keeping computational complexity manageable through focused analysis of specific data portions.
Data Source
AI summary
Systems and methods are disclosed herein for performing unstructured lane estimation. In one example, unstructured lane estimation involves the steps of receiving an overhead representation of a region and a rasterized trace generated from sensors of a vehicle traveling in the region and outputting, using a machine learning model, probabilities that neighboring pixels of pixels forming the rasterized trace is part of a lane based on the overhead representation and the rasterized trace.