YOLO-Based Running Record System
Automated Military Fitness Measurement
Overview
Developed through a 2024 military–academic AI program, this project explored how computer vision could reduce the personnel required for military fitness tests.
The project was developed by a collaborative team consisting of professors from Kyonggi University, military officers and NCOs from the ROK-DCC, and me as an enlisted soldier. The system uses YOLO to recognize participants’ bib numbers and count completed laps from video.
My work focused on improving recognition reliability through parameter tuning and a revised approach to aggregating recognition results.
Problem
During running tests at my unit, each participant was paired with a person responsible for counting laps. This one-to-one arrangement required substantial manpower and made test administration difficult when personnel were limited.
The project aimed to automate lap counting from camera footage so that fewer people would be needed for manual monitoring.
Reliable identification was a key technical challenge. As participants ran, folds in clothing and arm movements could obscure their bib numbers, interrupting recognition and making continuous tracking unreliable.
My role
- Tuned YOLO parameters to improve bib-number recognition reliability under running conditions.
- Identified sections of the running route where bib numbers were most consistently recognizable.
- Analyzed recognition failures caused by clothing folds and arm occlusion.
- Proposed a section-based aggregation approach that accumulates recognition results in designated areas rather than relying on continuous tracking throughout the run.
Approach
- The initial design aimed to track runners continuously in real time.
- However, bib numbers were not consistently visible during running, so individual recognition results could be missing or unreliable.
- To address this, I tuned the detection parameters and identified sections where recognition was more stable.
- I then proposed concentrating recognition on those sections and aggregating observations across multiple frames.
- This approach was intended to support reliable participant identification and lap counting despite intermittent bib-number visibility.
System architecture
- Video input: Camera footage of participants during the running test.
- Participant recognition: YOLO-based bib-number recognition to identify individual runners.
- Proposed observation aggregation: Recognition results accumulated across frames within designated sections to reduce the impact of intermittent occlusion.
- Lap-count output: The number of completed laps for each participant.
Technologies
Results
- The related program received an Excellence Award (IITP President's Award).