DEJAY ONE AI LAB

Our Lab

We develop and commercialize advanced AI solutions powered by world-class Vision AI and multimodal LLMs.

Contact Us Explore Our AI
VISION AI

2026 Global Leaderboard #1 · Vision AI

Our technology continuously tracks the same person across dense crowds and challenging capture conditions.
CrowdBoost, our proprietary Vision AI engine, achieved #1 rankings on both MOT17 and MOT20.

MOT17
Robust Pedestrian Tracking

Uninterrupted tracking has been proven across moving-camera, low-resolution, and low-light footage.

Moving-Camera Footage
Low-Resolution Footage
Low-Light Footage
MOT20
High-Density Crowd Tracking

Multi-object movement flows are tracked accurately in dense settings such as train stations, festivals, and stadiums.

Station CCTV
Festival CCTV
Stadium CCTV
01

Multi-Camera Tracking

Links movement data for the same person across different cameras into one continuous path. Tracks key entry points, cross-zone movement, and repeated routes.

Multi-Camera Tracking
02

3D Spatial Mapping

Aligns camera locations and fields of view with real-world 3D coordinates and synchronizes time data. Integrates existing CCTV feeds by floor and zone for a complete view of movement.

3D Spatial Mapping
03

Object Detection

Reliably detects pedestrians and objects in dense, complex footage. Precisely extracts location and movement data.

Object Detection
04

Movement Analytics

Analyzes entries, exits, average dwell time, pass-through rates, movement frequency, and congestion by time of day. Heatmaps and before-and-after comparisons make key operational metrics easy to understand.

Movement Analytics
05

Personal Data De-identification

De-identifies sensitive personal data, including faces, and analyzes only anonymized movement paths and usage patterns. Data scope, retention periods, and access controls can be configured to match internal policies and applicable laws. Insights are delivered through real-time dashboards and scheduled reports.

Personal Data De-identification
LLM · RAG

Advanced LLMs with Multimodal Reasoning

Up-to-date knowledge, multimodal analysis across text and images, and evidence-based reasoning enable accurate AI decisions—even in complex contexts.

Intelligent Vision AI × LLM Analysis
Fast, Accurate AI-Powered Review
01
Vision-Language Extraction

Precisely extracts the text and visual information required for analysis.

02
RAG-Based Regulatory Cross-Check

Retrieves and cross-checks evidence against continuously updated notices and regulations.

03
Contextual LLM Reasoning

Determines whether differently worded statements are equivalent in meaning.

04
Evidence Citation

Provides the supporting regulations and sources for reliable review.

05
Automated Update Pipeline

Automatically updates the database when notices or laws are revised.

06
Parallel Multi-Country Review

Cross-checks regulations across multiple countries in parallel.

07
Low-Latency Real-Time Inference

Processes high review volumes without delay for efficient large-scale review.

PATENT

Registered Patents

The Vision AI, VLM, and smart factory technologies developed at the DEJAY ONE AI Lab are protected by patents registered with the Korean Intellectual Property Office.

10 patents
Patent certificate No. 10-3042943
Patent No. 10-3042943
Method for verifying the consistency of vision-language model-based safety judgments and generating traceable evidence packages
Registered
Oct 30, 2026
Patent certificate No. 10-3038197
Patent No. 10-3038197
Method and apparatus for deriving real-world coordinates and semantic paths from coordinate-based object trajectories in CCTV video
Registered
Oct 21, 2026
Patent certificate No. 10-3038196
Patent No. 10-3038196
Method for tracking objects in multi-camera images and electronic device for performing the same
Registered
Oct 21, 2026
Patent certificate No. 10-3038195
Patent No. 10-3038195
Method for re-identifying occluded lost object and restoring movement trajectory, and electronic device performing the same
Registered
Oct 21, 2026
Patent certificate No. 10-3033882
Patent No. 10-3033882
Method for operating an intelligent work violation detection apparatus based on standard operating procedures (SOP), work permits, and real-time retrieval-augmented generation
Registered
Oct 13, 2026
Patent certificate No. 10-3027919
Patent No. 10-3027919
System for safety situation judgment based on privacy-preserving de-identified scene graph and vision-language model (VLM)
Registered
Sep 29, 2026
Patent certificate No. 10-3025488
Patent No. 10-3025488
System for de-identified textualization and dynamic prioritization of public safety events via semantic analysis of multi-source video data
Registered
Sep 22, 2026
Patent certificate No. 10-3013712
Patent No. 10-3013712
Variable Micro-VLM control method based on event triggering of an object detector
Registered
Aug 31, 2026
Patent certificate No. 10-2694245
Patent No. 10-2694245
Indoor location-based smart factory system using beacon
Registered
Aug 7, 2024
Patent certificate No. 10-2636781
Patent No. 10-2636781
Smart factory system for integrated manufacturing operation management
Registered
Feb 8, 2024
PUBLICATIONS

Research Publications

Our research on multi-object tracking in dense crowds has been accepted for a long oral presentation at the ACCV 2026 HUFPEI Workshop.

2026
ACCV LONG ORAL

Coordinates, Not Candidates: Dissecting Detector Ensembling for Dense-Crowd Tracking

Doohyun Kim, Yonghoon Lee, Chaeyeon Cho, Sanghyun Kim · DEJAY

We analyzed detector ensembling for dense-crowd tracking by separating its effects on box coordinates, candidates, and scores, and showed that refining the coordinates alone improves performance. Applying this approach raised HOTA on the MOT20 benchmark from 67.98 to 69.16.

MOT20 HOTA 67.98 → 69.16
Multi-object trackingCrowd analysisDetector ensemblesBox votingEvaluation protocol
2026 #1 - MOT17 & MOT20 Leaderboards · Vision AI

DEJAY ONE

Transforming spaces into business growth.

VISION AI LLM VLM RAG Fine Tuning
Contact Us