Analysis and Development of Video-Based Unknown Object Recognition Technology (영상정보 기반의 미확인 객체 인식 기술 분석 및 개발 용역)

Object detection, classification and localization in port CCTV.

Visual Intelligence & Logistics

2020-11-09 – 2020-12-24 · Completed

Korea Railroad Research Institute (KRRI)

Results

  • Separated detection and similarity matching to extend recognition through reference images.
  • Combined semi-automatic data generation and camera calibration to produce spatial coordinates.

Evaluation on held-out CCTV footage.

Conditions Object recognition Classification success
Rainy night Approximately 95% Approximately 90%
Rainy day Approximately 95% Approximately 95%
Daytime Approximately 97% Approximately 97%

At the designated test points, mean localization error was 0.1697 m and maximum error was 0.6022 m. Average processing rates were 20 fps on a desktop and 5 fps on an embedded PC.