Neural Sensor Fusion for Dense Occupancy Grid Estimation

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing sensor systems in advanced driver assistance systems face challenges in generating dense environmental information efficiently, as low-cost sensors provide sparse data while high-cost sensors like LIDAR are not suitable for mass market applications, and smart sensors lack raw image data processing.

Innovation Solution

A neural network device and method for sensor fusion that combines feature-level data from multiple sensors to generate a dense environmental model, using unsupervised training to process sparse input data and classify large unobserved regions without requiring continuous information gathering.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If LIDAR sensors are used to provide dense environmental information, then measurement precision is improved, but device cost increases

Engineering Contradiction:
Improveenvironmental information densityVSAvoiddevice cost
Core Design Contradiction:
Measurement precisionVSEase of manufacture

Solution Approach 1:

The patent creates a virtual copy of dense LIDAR environmental data by training a neural network on LIDAR data during an offline phase. The trained network then generates synthetic dense environmental information from sparse radar and camera inputs, effectively copying the output characteristics of LIDAR without requiring physical LIDAR hardware in the deployment system.

Inventive Principle:
Principle #26Copying

Solution Approach 2:

The patent replaces expensive LIDAR sensors with cheaper radar and camera sensors that provide sparse data. By using cost-effective sensors and compensating through neural network processing, the system achieves dense environmental modeling capability without the high hardware cost of LIDAR.

Inventive Principle:
Principle #27Cheap short-living objects (Disposable)

2Measurement precision

If data is gathered over time to obtain meaningful environmental model, then measurement precision is improved, but productivity decreases

Engineering Contradiction:
Improveenvironmental model qualityVSAvoiddata processing speed
Core Design Contradiction:
Measurement precisionVSProductivity

Solution Approach 1:

The patent performs preliminary training of the neural network offline using dense LIDAR data and sparse radar/camera data pairs. This pre-computed knowledge is stored in the network weights, enabling the system to generate dense environmental models instantly during runtime without requiring time-consuming data gathering or processing at that moment.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent replaces the mechanical approach of gathering data over time with multiple sensors with a computational approach using a trained neural network. The network directly transforms sparse sensor inputs into dense environmental models through learned mappings, eliminating the need for temporal data accumulation.

Inventive Principle:
Principle #28Mechanics substitution (Replace mechanical system)

3Device complexity

If smart sensors are used to process detected data within the sensor, then device complexity is reduced, but loss of information increases

Engineering Contradiction:
Improvesensor processing architectureVSAvoidraw image data
Core Design Contradiction:
Device complexityVSLoss of information

Solution Approach 1:

The patent introduces a neural network as an intermediary processing layer between the smart sensors and the environmental model generation. The network receives processed data from smart sensors and reconstructs dense environmental information that would otherwise be lost, acting as a mediator that recovers information rather than requiring access to原始 raw data.

Inventive Principle:
Principle #24Intermediary (Mediator)

Data Source

PatentEP3745299B1Neural network device and method using a neural network for sensor fusion
Publication Date: 2026.02.25 INFINEON TECHNOLOGIES AG
  • EP3745299B1 patent drawingFigure 1A
  • EP3745299B1 patent drawingFigure 1B
  • EP3745299B1 patent drawingFigure 1C

AI summary

A neural network device and a method using a neural network for sensor fusion are disclosed, wherein the neural network device includes a neural network (136) and wherein the neural network (136) is configured to: process a first grid (132) including a plurality of grid cells, wherein the first grid (132) represents at least a first portion of a field of view (108) of a first sensor (104), wherein at least one grid cell has information about an occupancy of the first portion of the field of view (108) assigned to the at least one grid cell, the information being based on data provided by the first sensor (104); process a second grid (134) including a plurality of grid cells, wherein the second grid (134) represents at least a second portion of a field of view (110) of a second sensor (106), wherein at least one grid cell has information about an occupancy of the second portion of the field of view (110) assigned to the at least one grid cell, the information being based on data provided by the second sensor (106); and fuse the processed first grid (132) with the processed second grid (134) into a fused grid (138), wherein the fused grid (138) includes information about the occupancy of the first portion of the field of view (108) of the first sensor (104) and the occupancy of the second portion of the field of view (110) of the second sensor (106).