Sports Video Object Detection With Dewarped Square Views

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing sports video production systems struggle with inaccurate detection and tracking of small, fast-moving objects like a ball due to image deformation and inefficient algorithms, leading to false positives and high resource consumption.

Innovation Solution

A method using synchronized physical cameras to create panoramic video streams, de-warping images into optimized square views, and employing a deep learning neural network like YOLO for efficient object detection, allowing for accurate and resource-efficient tracking of objects.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If image recognition techniques with background subtraction are used to detect moving objects, then the system can identify moving parts, but it produces false positives for small objects like a ball

Engineering Contradiction:
Improvedetection accuracyVSAvoidfalse positive rate
Core Design Contradiction:
Measurement precisionVSReliability

Solution Approach 1:

The system segments the detection task by dividing the image processing into multiple specialized stages: panoramic stitching from multiple cameras, deformation correction, and then detection. This segmentation allows each stage to optimize for its specific function, improving overall detection accuracy while reducing false positives for small objects like balls.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent introduces an intermediary processing step that corrects image deformation before detection. This intermediary correction layer acts as a mediator between the raw camera input and the detection algorithm, transforming distorted images into geometrically accurate representations that enable reliable ball detection without false positives.

Inventive Principle:
Principle #24Intermediary (Mediator)

2Loss of information

If physical cameras record images for detection, then the system can capture real-world scenes, but the images are already deformed representations that reduce detection accuracy

Engineering Contradiction:
Improveimage deformationVSAvoiddetection accuracy
Core Design Contradiction:
Loss of informationVSMeasurement precision

Solution Approach 1:

Instead of accepting deformed images as-is for detection, the system inverts the problem by first identifying and correcting the deformation. The patent applies geometric correction transformations that reverse the camera-induced distortion, converting deformed panoramic images back into accurate geometric representations suitable for precise object detection.

Inventive Principle:
Principle #13The other way round (Inversion)

3Measurement precision

If detection algorithms process all video frames to track objects, then complete coverage is achieved, but memory use and calculation capacity requirements increase significantly

Engineering Contradiction:
Improvetracking accuracyVSAvoidcomputational resources
Core Design Contradiction:
Measurement precisionVSUse of energy by moving object

Solution Approach 1:

The system performs preliminary actions by pre-stitching panoramic views from multiple cameras and pre-correcting geometric deformations before the detection phase. This preliminary processing creates optimized input data that reduces the computational burden during actual object tracking, allowing accurate detection with lower memory usage and calculation requirements.

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP4172937B1Method for detecting and/or tracking moving objects within a certain zone and sports video production system in which such a method is implemented
Publication Date: 2025.09.17 MOBILE VIEWPOINT BV
  • EP4172937B1 patent drawingFigure 1
  • EP4172937B1 patent drawingFigure 1
  • EP4172937B1 patent drawingFigure 2

AI summary

Method for detecting moving objects (10-12) within a certain zone (5), comprising the steps of: - providing multiple physical cameras (4)) around the zone (5); - composing a panoramic video stream; - defining one or more virtual camera view(s) (25) by selecting corresponding partial views (26) of the panoramic views (23) and by de-warping the selected, partial views (26) into square views (27);10 - feeding each de-warped view (27) to a deep learning neural network; and, - performing a detection in the corresponding part of the zone (5).