Neural Network Classification and Re-identification Architecture

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Surveillance systems face computational intensity in accurately tracking individuals and objects across multiple cameras, and the training of neural networks for re-identification is resource-intensive due to the need for both classification and re-identification processes.

Innovation Solution

A neural network architecture that performs feature extraction and provides features to both classification and re-identification layers in parallel, reducing the need for repeated feature extraction processes by overlapping these processes during the training phase.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If the neural network performs both classification and re-identification by searching through all images, then the accuracy of tracking targets is improved, but the computational resources and time required increase significantly

Engineering Contradiction:
Improvetracking accuracyVSAvoidprocessing efficiency
Core Design Contradiction:
Measurement precisionVSProductivity

Solution Approach 1:

The patent segments the neural network into two distinct branches: a classification branch that categorizes objects into predefined classes, and a re-identification branch that performs fine-grained discrimination between similar instances. This segmentation allows each branch to specialize in its specific task, improving overall tracking accuracy while reducing the computational burden compared to a single monolithic network searching through all images.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent extracts and separates the re-identification function from the general classification process. By taking out the re-identification branch as an independent component with its own feature extraction and comparison mechanisms, the system can focus computational resources on specific discrimination tasks rather than searching through all images for every target identification.

Inventive Principle:
Principle #2Taking out (Extraction)

2Device complexity

If the system uses a single neural network for both classification and re-identification, then the device complexity is reduced, but the training resource allocation becomes inefficient

Engineering Contradiction:
Improvenetwork structureVSAvoidtraining efficiency
Core Design Contradiction:
Device complexityVSEase of manufacture

Solution Approach 1:

The neural network is segmented into two parallel branches with distinct architectures optimized for their specific functions. The classification branch uses standard classification layers, while the re-identification branch employs specialized feature extraction and metric learning components. This segmentation enables independent training and optimization of each branch, improving training efficiency while maintaining manageable system complexity.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent implements a shared feature extraction backbone that serves both classification and re-identification tasks. This multi-functional base network extracts general object features that are then processed by task-specific branches, allowing the system to maintain relatively low overall complexity while enabling efficient specialized training for each function through the shared representations.

Inventive Principle:
Principle #6Universality (Multi-functionality)

Data Source

PatentEP3800577A1Classification and re-identification using a neural network
Publication Date: 2021.04.07 JOHNSON CONTROLS TYCO IP HLDG LLP
  • EP3800577A1 patent drawingFigure 1
  • EP3800577A1 patent drawingFigure 2
  • EP3800577A1 patent drawingFigure 3

AI summary

Aspects of the present disclosure include methods, systems, and non-transitory computer readable media that perform the steps of receiving one or more snapshots, extracting one or more features from the one or more snapshots, and providing the one or more features to a first classification layer for classifying a first target and a second classification layer for re-identifying a second target.