Perception Data Fusion for Reliable Autonomous Obstacle Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine perception systems struggle to accurately combine the strengths of different sensor modalities and processing pipelines, leading to inaccuracies in obstacle detection and decision-making in complex environments.
Innovation Solution
Fusing information generated by learned models (e.g., deep neural networks) with non-learned processes (e.g., algorithmic methods) to refine and improve data, leveraging the strengths of both approaches for robust obstacle detection in near-range environments.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If learned models are used for obstacle detection, then the ability to identify unclassified objects and estimate object shapes and sizes is improved, but performance in edge cases where training data is limited deteriorates
Solution Approach 1:
The patent combines learned models (deep neural networks) with non-learned processes (algorithmic methods) into a unified perception system. The learned models process sensor data to generate first information about objects, while non-learned processes generate second information using classical computer vision techniques. These two information streams are then fused to produce third information that leverages the strengths of both approaches, resolving the contradiction between detection accuracy and reliability in edge cases.
Solution Approach 2:
The perception system uses a composite processing pipeline that integrates two different processing paradigms: learned models and non-learned algorithms. This composite approach allows the system to benefit from the pattern recognition capabilities of neural networks while simultaneously utilizing the interpretability and reliability of classical algorithms, particularly in scenarios where training data may be insufficient.
2Reliability
If multiple sensor modalities are combined for obstacle detection, then the robustness of detection is improved, but the complexity of the perception system increases
Solution Approach 1:
The patent segments the perception system into distinct processing pipelines: one for learned models and another for non-learned processes. Each pipeline processes sensor data independently through its specialized methodology, and the results are subsequently fused. This segmentation allows the system to handle multiple sensor modalities (image, RADAR, LiDAR, ultrasonic) without creating a monolithic complex structure, as each sensor type can be processed by the most appropriate pipeline.
Solution Approach 2:
The fusion component serves multiple functions: it integrates outputs from both learned and non-learned pipelines, reconciles conflicting detections, fills in missing information, and produces a unified perception result. This multi-functional design allows the system to process diverse sensor modalities through a single integration point, reducing overall system complexity while maintaining robust detection capabilities.
Data Source
AI summary
Techniques for fusing first information generated using one or more learned models with second information generated using one or more non-learned processes to generate third information including one or more updated versions of the first information and/or the second information. In some examples, the first information may indicate one or more locations associated with one or more first objects in an environment, and the second information may indicate one or more attributes associated with one or more second objects in the environment. In some instances, the learned model(s) may generate the first information based at least on first sensor data generated using one or more first sensors of a machine, and the non-learned process(es) may generate the second information based at least on second sensor data generated using one or more second sensors of the machine.