Vehicle Sensor Fusion Using Sensor-Specific Pose Probability Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing sensor data fusion methods for autonomous vehicles often fail to achieve optimal precision, accuracy, and coverage due to processing sensor measurements independently, leading to inefficiencies in understanding the surrounding environment.
Innovation Solution
A system that fuses object pose probability data from multiple sensors, such as radar and stereo cameras, using sensor-specific models to combine data at an early abstraction level, allowing for improved object detection and tracking.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If sensor measurements are processed independently per sensor, then device complexity is reduced and ease of operation is improved, but measurement precision and reliability of object detection deteriorate
Solution Approach 1:
The patent segments the fusion process into distinct stages: low-level feature fusion (combining raw sensor measurements), mid-level fusion (combining detected objects and features), and high-level fusion (combining semantic information). This segmentation allows the system to achieve high measurement precision through multi-level fusion while managing complexity by processing different types of data at appropriate granularities.
Solution Approach 2:
The patent introduces a temporal dimension to the fusion process by performing both spatial fusion (across sensors) and temporal fusion (across time steps). This multi-dimensional approach enhances detection precision by considering both spatial relationships between sensors and temporal evolution of detected objects, while the structured dimensional organization helps manage overall system complexity.
2Measurement precision
If multiple sensors are fused at high-level representation, then device complexity is reduced, but measurement precision and coverage requirements are insufficiently met
Solution Approach 1:
The patent divides the fusion architecture into three distinct levels: low-level fusion for raw measurements, mid-level fusion for detected objects, and high-level fusion for semantic interpretation. This segmentation enables the system to meet high precision requirements by fusing data at multiple granularities while managing architectural complexity through clear separation of concerns at each level.
Solution Approach 2:
The patent performs preliminary processing and feature extraction at each sensor level before fusion, preparing data in advance for efficient combination. This preliminary action at the low and mid levels ensures that when high-level fusion occurs, the data is already optimized, thereby achieving high environmental perception accuracy while reducing the computational complexity during the actual fusion operation.
3Productivity
If sensor data is processed independently, then computational overhead is reduced, but coverage and accuracy of object detection deteriorate
Solution Approach 1:
The patent segments processing tasks by assigning different computational responsibilities to different fusion levels: low-level fusion handles efficient raw data combination, mid-level fusion processes object detections, and high-level fusion performs semantic analysis. This segmentation improves detection accuracy through comprehensive fusion while maintaining processing efficiency by optimizing computational operations at each level.
Solution Approach 2:
The patent applies partial fusion strategies where only relevant sensor data and features are combined at each level rather than processing all data exhaustively. This partial action approach maintains high detection accuracy by focusing computational resources on critical data while improving overall processing efficiency by avoiding unnecessary computations.
Data Source
AI summary
An information processing device for a vehicle for sensor data fusion for object detection, including circuitry configured to:obtain, based on obtained first sensor data from a first sensor of the vehicle and a first predetermined object pose probability model, first object pose probability data, wherein the first predetermined object pose probability model is specific for the first sensor;obtain, based on obtained second sensor data from a second sensor of the vehicle and a second predetermined object pose probability model, second object pose probability data, wherein the second predetermined object pose probability model is specific for the second sensor; andfuse the first and the second object pose probability data to obtain fused object pose probability data for object detection.