Autonomous Robot Occupancy Grid Segmentation With Spectral Neural Networks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Occupancy grids in autonomous vehicles do not provide information on the nature and geometry of obstacles, and existing methods for real-time obstacle recognition on embedded systems with limited computing resources are computationally expensive and inefficient.
Innovation Solution
A method using dense neural networks in the frequency domain to segment material objects by transforming occupancy grids into spectral representations, applying dimension reduction and executing an AI algorithm to identify vehicles or objects, reducing computational cost and memory footprint.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If occupancy grids are used to represent the environment, then sensor data fusion is achieved, but information on obstacle nature and geometry is lost
Solution Approach 1:
The occupancy grid is segmented into multiple semantic categories (e.g., vehicle, pedestrian, cyclist, animal, unknown) rather than treating all occupied cells uniformly. This segmentation preserves obstacle nature information while maintaining the fusion capability of occupancy grids.
Solution Approach 2:
The patent extends the traditional 2D occupancy grid by adding a semantic dimension, transforming it from a binary occupied/not-occupied representation to a multi-class semantic map. This dimensional extension preserves geometric information while enriching it with obstacle nature classification.
2Measurement precision
If traditional obstacle recognition methods are used on embedded systems, then obstacle detection is achieved, but computational cost is too high for real-time execution
Solution Approach 1:
The patent changes the parameter representation from raw sensor data or high-resolution images to occupancy grid probabilities. This parameter transformation reduces data dimensionality and computational complexity while maintaining detection accuracy, enabling real-time execution on embedded systems.
Solution Approach 2:
The patent replaces complex mechanical vision processing systems with a probabilistic occupancy grid-based approach. This substitution uses statistical methods and probability theory instead of traditional image processing mechanics, reducing computational burden while maintaining detection precision.
3Productivity
If high computing power is provided for real-time perception, then processing speed is improved, but hardware constraints of autonomous vehicles are violated
Solution Approach 1:
The patent extracts only the essential information needed for obstacle detection from raw sensor data, representing it in a compact occupancy grid format. This extraction removes redundant data while preserving critical perception capabilities, reducing hardware requirements.
Solution Approach 2:
The patent processes only the necessary portion of sensor data at full resolution, using occupancy grids to represent the environment at an appropriate level of detail. This partial processing approach achieves real-time performance without requiring excessive computing power.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
The present invention relates to a method and an associated device for segmenting material objects in a space surrounding an autonomous robot. In particular, the invention addresses a method based on the analysis of frequency spectra of occupancy grids by dense neural networks to segment vehicles in the environment of an autonomous vehicle.