Autonomous Vehicle Distance Prediction Using Curved Ground Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional systems for detecting distances to objects or obstacles in autonomous vehicles rely on the assumption of a flat ground plane, leading to inaccurate predictions when the road surface is curved, and are also affected by variations in camera parameters during training and deployment, resulting in safety and performance issues.
Innovation Solution
A deep neural network (DNN) is trained using sensor data from LIDAR, RADAR, and other depth sensors to predict distances, leveraging ground truth data for accurate image-based predictions, and a camera adaptation algorithm is used to account for camera model variations, with safety bounds computation to ensure accurate and reliable distance estimation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If conventional systems assume a flat ground plane for distance calculation, then the system complexity is reduced and calculation is simplified, but measurement precision deteriorates when the road surface is curved
Solution Approach 1:
The patent applies curvature by modeling the ground plane as a curved surface rather than a flat plane. The system estimates road curvature parameters and uses them to adjust the ground plane model, allowing accurate distance calculations on curved road surfaces while maintaining computational efficiency through parametric modeling.
Solution Approach 2:
The system changes parameters by introducing curvature parameters (such as curvature radius and curvature angle) to the ground plane model. These parameters are estimated from sensor data and used to dynamically adjust the distance calculation model, enabling adaptation to different road conditions without increasing system complexity.
2Measurement precision
If a deep neural network is trained with multi-sensor data (LIDAR, RADAR, depth sensors), then measurement precision improves for distance prediction, but device complexity and training time increase
Solution Approach 1:
The patent merges multiple sensor data types (LIDAR, RADAR, depth sensors, and image data) into a unified training framework. By combining these diverse data sources, the system achieves robust distance predictions that are insensitive to individual sensor variations and work accurately across different road conditions and camera models.
Solution Approach 2:
The system introduces an intermediary ground truth generation process that uses depth sensors and LIDAR/RADAR data to create accurate distance labels for training the image-based DNN. This intermediary process bridges the gap between raw sensor data and the final distance prediction task, enabling precise training without requiring complex direct multi-sensor fusion during inference.
3Measurement precision
If manual annotation is used for training data preparation, then measurement precision can be ensured, but productivity decreases due to time-consuming annotation process
Solution Approach 1:
The system implements self-service by using depth sensors, LIDAR, and RADAR to automatically generate ground truth distance data for training the DNN. This automated ground truth generation eliminates the need for manual annotation while maintaining high accuracy, as the physical sensors provide direct measurements of distance that serve as reliable training labels.
Solution Approach 2:
The patent replaces the mechanical manual annotation process with an automated sensor-based measurement system. Instead of human annotators manually measuring and labeling distances in images, the system uses physical sensors (LIDAR, RADAR, depth sensors) to automatically capture accurate distance data, substituting human labor with automated sensing and processing.
Data Source
AI summary
In various examples, a deep neural network (DNN) is trained—using image data alone—to accurately predict distances to objects, obstacles, and/or a detected free-space boundary. The DNN may be trained with ground truth data that is generated using sensor data representative of motion of an ego-vehicle and/or sensor data from any number of depth predicting sensors—such as, without limitation, RADAR sensors, LIDAR sensors, and/or SONAR sensors. The DNN may be trained using two or more loss functions each corresponding to a particular portion of the environment that depth is predicted for, such that—in deployment—more accurate depth estimates for objects, obstacles, and/or the detected free-space boundary are computed by the DNN. In some embodiments, a sampling algorithm may be used to sample depth values corresponding to an input resolution of the DNN from a predicted depth map of the DNN at an output resolution of the DNN.