Multiscale Neural Network Crowd Density Estimation

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current crowd density estimation methods, particularly those using neural networks, face challenges in accuracy due to obstacles and significant depth of field, leading to unreliable people counting in dense crowds.

Innovation Solution

A multiscale neural network approach that processes input images and identification masks from mobile communicating devices, using multiple convolutional processing layers to generate a density map, improving accuracy by considering different scales and spatial coordinates.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If traditional crowd counting methods (manual or computer-assisted) are used, then the counting process can be performed, but the counting time is slow and the error rate is high

Engineering Contradiction:
Improvecounting accuracyVSAvoidcounting time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The patent replaces manual or computer-assisted mechanical counting methods with an automated optical system using a camera and neural network algorithm. The camera captures images of the crowd, and the neural network automatically processes these images to count individuals, eliminating the need for manual counting while significantly improving both speed and accuracy.

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

2Measurement precision

If laser counting by helicopter is used, then precise counting can be achieved, but the device complexity increases and setup becomes difficult

Engineering Contradiction:
Improvecounting precisionVSAvoidsystem complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent extracts the essential counting function from complex aerial laser systems and implements it using a simple ground-based camera. By removing the need for helicopters, qualified personnel, and complex laser equipment, the system achieves comparable precision through a much simpler setup that can be deployed easily in various environments.

Inventive Principle:
Principle #2Taking out (Extraction)

Solution Approach 2:

The patent uses a camera to capture visual information of the crowd, creating an image copy that can be analyzed by a neural network. This approach replicates the counting capability of complex laser systems but uses optical copying through imaging, which is simpler, safer, and easier to implement.

Inventive Principle:
Principle #26Copying

3Productivity

If existing neural network methods are used for crowd counting, then automated counting can be performed, but accuracy decreases in dense crowds or when people are hidden behind obstacles

Engineering Contradiction:
Improveautomated counting capabilityVSAvoiddetection accuracy
Core Design Contradiction:
ProductivityVSMeasurement precision

Solution Approach 1:

The patent addresses the limitation of 2D image-based counting by integrating 3D depth information from the mobile device's depth sensor. This additional dimensional data allows the neural network to better distinguish individuals in dense crowds and detect people partially hidden behind obstacles, significantly improving accuracy while maintaining automated processing capability.

Inventive Principle:
Principle #17Another dimension (Dimensionality change)

Data Source

PatentEP4033399B1Computer device and method for estimating the density of a crowd
Publication Date: 2023.10.18 BULL SA
  • EP4033399B1 patent drawingFigure 1
  • EP4033399B1 patent drawingFigure 2
  • EP4033399B1 patent drawingFigure 3

AI summary

The invention relates to a computer device and a method for estimating the density of a crowd (10) using mobile communicating devices (30), said method comprising the use of a trained multiscale neural network, said neural network comprising at least two sub-networks of neurons, said method comprising a fusion (400), by the trained multiscale neural network, of at least one matrix (Mx1, Mx2, Mx3) corresponding to the processing of the input image (21) and of at least one matrix (Mx5) corresponding to the processing of the identification data of communicating devices (30) so as to obtain a density map (Dm) of the crowd (10) whose density is to be estimated.