Drivable Area Detection for Redundant Autonomous Navigation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing autonomous vehicle navigation systems rely on singular positioning methods, creating a single point of failure for safety-critical subsystems, and lack robustness in detecting drivable areas without relying on map-based localization.
Innovation Solution
A method utilizing machine learning for drivable area detection through pixel-based image segmentation of camera data, combined with LiDAR point cloud projection and odometry information, to generate real-time drivable area maps independent of traditional localization systems.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If map-based localization is used for navigation, then the vehicle can determine its position using existing infrastructure, but the system creates a single point of failure for safety-critical subsystems
Solution Approach 1:
The patent segments the positioning function into multiple independent components: map-based localization subsystem and sensor-based drivable area detection subsystem. Each subsystem operates independently and can provide navigation information, eliminating the single point of failure while distributing the positioning complexity across separate modular components.
Solution Approach 2:
The patent changes the fundamental parameter of position determination from relying solely on map coordinates to using direct sensor measurements of drivable areas. This parameter change allows the system to bypass localization failures by directly detecting navigable spaces through LiDAR and camera data, improving reliability without adding complex integration layers.
2Reliability
If sensor-based drivable area detection is implemented, then the system provides redundant navigation capability, but the computational overhead increases
Solution Approach 1:
The patent applies preliminary action by pre-training machine learning models offline to recognize drivable area patterns. During runtime, the system only needs to perform inference rather than full training, significantly reducing computational energy consumption while maintaining high reliability through the redundant detection capability.
Solution Approach 2:
The patent substitutes complex mechanical computation and manual path planning with machine learning-based automated detection. The ML models automatically identify drivable areas from sensor data, replacing energy-intensive traditional algorithms with optimized neural network inference that requires less computational power while providing redundant navigation information.
3Productivity
If pixel-based image segmentation is used for drivable area detection, then the system achieves real-time processing, but the measurement precision requirements increase
Solution Approach 1:
The patent merges multiple sensor data sources (LiDAR point clouds and camera images) to compensate for the precision limitations of individual sensors. The LiDAR provides accurate depth information while the camera provides high-resolution texture data, and their fusion through pixel-based segmentation achieves both real-time processing speed and precise drivable area boundary detection.
Solution Approach 2:
The patent introduces machine learning models as intermediaries between raw sensor data and drivable area determination. These ML models process the sensor inputs through learned feature representations, enabling real-time inference while achieving high measurement precision for drivable area boundaries that would be difficult to obtain through traditional image processing alone.
Applied Scientific Principles
This section explains which scientific principles are used to turn an abstract innovation direction into a practical engineering solution.
Function Achieved in This Case
Provides a redundant navigation system that allows autonomous vehicles to continue operating safely within detected drivable areas, even in the event of localization or perception failures, with minimal additional computational overhead.
Implementation Method 1
receiving, with at least one processor, a point cloud from a depth sensor
Implementation Method 2
receiving, with the at least one processor, image data from a camera
Data Source
AI summary
Enclosed are embodiments for navigation with drivable area detection. In an embodiment, a method comprises: receiving a point cloud from a depth sensor, receiving image data from a camera; predicting at least one label indicating a drivable area by applying machine learning to the image data; labeling the point cloud using the at least one label; obtaining odometry information; generating a drivable area by registering the labeled point cloud and odometry information to a reference coordinate system; and controlling the vehicle to drive within the drivable area.