Distributed Neural Network for Drone Object Identification

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Convolutional neural networks for real-time object identification in high-resolution videos require significant computer resources, exceeding the size, weight, and power constraints of small and medium-sized remote-controlled vehicles, and are costly, with limited ability to continue learning independently.

Innovation Solution

A distributed neural network system where the resource-intensive portion is moved to a base station, with only feature maps or characterization processed on the vehicle, allowing for real-time object identification and continuous training by dividing the network into a less performant front-end and a highly performant base-station side, enabling efficient video processing and transmission.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If a highly performant neural network is deployed on the vehicle for real-time object identification, then object identification accuracy is improved, but size, weight and power consumption increase beyond acceptable limits

Engineering Contradiction:

Improveobject identification accuracyVSAvoidpower consumption

Core Design Contradiction:

Measurement precisionVSUse of energy by moving object

Solution Approach 1:

The neural network is segmented into two parts: a lightweight network on the vehicle for real-time processing and a deep network on the base station for accurate identification. The vehicle's neural network performs initial object area detection, while the base station's deep neural network performs final accurate identification, dividing the computational load to balance accuracy and power consumption.

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

A base station is introduced as an intermediary between the vehicle and the object identification process. The base station receives video data from the vehicle, performs deep neural network processing, and returns identification results, allowing the vehicle to maintain a lightweight network while achieving high accuracy through the base station's computational power.

Inventive Principle:

Principle #24Intermediary (Mediator)

2Measurement precision

If a highly performant neural network is deployed on the vehicle, then object identification capability is improved, but device size and weight increase

Engineering Contradiction:

Improveobject identification accuracyVSAvoidvehicle weight

Core Design Contradiction:

Measurement precisionVSWeight of moving object

Solution Approach 1:

The neural network processing is segmented between vehicle and base station. The vehicle contains only a lightweight neural network for detecting object areas, while the deep neural network for accurate identification is located on the base station, significantly reducing the weight of the vehicle while maintaining high identification accuracy.

Inventive Principle:

Principle #1Segmentation

3Measurement precision

If a deep neural network is deployed on the vehicle, then object identification accuracy is improved, but computational resources and cost increase

Engineering Contradiction:

Improveobject identification accuracyVSAvoidcomputational resource requirements

Core Design Contradiction:

Measurement precisionVSDevice complexity

Solution Approach 1:

The deep neural network is segmented and deployed on the base station rather than the vehicle. The vehicle's lightweight network detects object areas, and the base station's deep network performs accurate identification, reducing computational resource requirements and cost on the vehicle while maintaining high accuracy.

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

The neural network processing is copied across two locations: a simplified version on the vehicle for real-time detection and a complete deep version on the base station for accurate identification. This copying strategy allows the vehicle to function with limited resources while the base station provides full computational capability.

Inventive Principle:

Principle #26Copying

4Measurement precision

If video data is transmitted from vehicle to base station for processing, then object identification accuracy is improved, but transmission bandwidth requirements increase

Engineering Contradiction:

Improveobject identification accuracyVSAvoiddata transmission volume

Core Design Contradiction:

Measurement precisionVSQuantity of substance

Solution Approach 1:

Only the necessary video data for object identification is extracted and transmitted to the base station, rather than transmitting all processed data. The vehicle's lightweight network pre-processes video to identify object areas, extracting only relevant regions for transmission to the base station, reducing overall data transmission volume.

Inventive Principle:

Principle #2Taking out (Extraction)

Data Source

PatentUS11816567B2System for identifying objects by means of distributed neural networks

Publication Date: 2023.11.14 ROCKWELL COLLINS DEUTLAND

AI summary

In a system for identifying objects by means of distributed neural networks a resource-intensive proportion of the neural network is provided at a base station (on the ground), whereas a less resource-intensive proportion of the neural network, in particular the remaining proportion of the neural network, is provided at a front end (for example remote-controlled vehicle). The technical complexity for the front-end side and for the transmission of data between the base station and front end can thus be reduced.