Crowd Object Detection Using 3D Geometric Occlusion Context

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing object detection systems struggle in crowded scenes due to occlusion and scale variations, leading to missed detections and merged classifications, as they primarily rely on appearance features without adequately incorporating geometric context.

Innovation Solution

Integrate geometric context into deep neural networks by using camera pose and scene geometry to model occlusion patterns, projecting regions of interest and occluders onto a fixed plane, and applying a geometric loss function to improve object detection accuracy.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If traditional object detection methods using region proposal and convolutional neural networks are used, then the system can identify objects based on appearance features, but detection accuracy deteriorates in crowded scenes due to occlusion and scale variations

Engineering Contradiction:
Improveobject detection accuracyVSAvoidocclusion and scale variations
Core Design Contradiction:
ReliabilityVSObject-affected harmful factors

Solution Approach 1:

The patent projects 2D image regions onto a 3D ground plane and back, adding geometric dimensionality to the detection process. This dimensional transformation allows the system to reason about occlusion and scale in 3D space, improving detection accuracy in crowded scenes where traditional 2D appearance-based methods fail.

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

Solution Approach 2:

The patent introduces a geometric context module as an intermediary between the image data and the classifier. This module projects regions of interest onto the ground plane, identifies potential occluders, and generates occluded RoIs, serving as a mediator that enriches the input features with geometric information before classification.

Inventive Principle:
Principle #24Intermediary (Mediator)

2Reliability

If the system processes all regions in the image data, then comprehensive object detection is achieved, but computational complexity increases due to the need to handle occlusion and geometric context

Engineering Contradiction:
Improvedetection completenessVSAvoidcomputational complexity
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The patent segments the image processing task into distinct components: region proposal, geometric projection onto ground plane, occluder identification, and occlusion zone generation. This segmentation allows each component to be optimized independently and processed efficiently, reducing overall computational complexity while maintaining detection completeness.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent performs preliminary geometric projection and occluder identification before the main classification step. By pre-processing the image data to generate occluded RoIs and their corresponding ground plane projections, the system prepares the data in advance, reducing the computational burden during the actual detection phase.

Inventive Principle:
Principle #10Preliminary action

3Measurement precision

If geometric context and ground plane projection are integrated into the neural network, then detection accuracy in crowded scenes improves, but the network architecture becomes more complex

Engineering Contradiction:
Improvelocation accuracyVSAvoidnetwork architecture complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent extracts geometric context information (camera pose, ground plane parameters, occlusion zones) as separate features and feeds them into the neural network alongside appearance features. By separating the geometric processing from the core neural network architecture, the system improves location accuracy without significantly complicating the network structure itself.

Inventive Principle:
Principle #2Taking out (Extraction)

Data Source

PatentEP3899786B1Detecting objects in crowds using geometric context
Publication Date: 2025.11.19 MICROSOFT TECHNOLOGY LICENSING LLC
  • EP3899786B1 patent drawingFigure 1
  • EP3899786B1 patent drawingFigure 2
  • EP3899786B1 patent drawingFigure 3

AI summary

A computing system is configured to train an object classifier. Monocular image data and ground-truth data are received for a scene. Geometric context is determined including a three-dimensional camera position relative to a fixed plane. Regions of interest (RoI) and a set of potential occluders are identified within the image data. For each potential occluder, an occlusion zone is projected onto the fixed plane in three-dimensions. A set of occluded RoIs on the fixed plane are generated for each occlusion zone. Each occluded RoI is projected back to the image data in two-dimensions. The classifier is trained by minimizing a loss function generated by inputting information regarding the RoIs and the occluded RoIs into the classifier, and by minimizing location errors of each RoI and each occluded RoI of the set on the fixed plane based on the ground-truth data. The trained classifier is then output for object detection.