Junhwan Kim

Junhwan Kim

Projects

Filters:

Real-time Object Detection Surveillance Software

A surveillance software that performs object detection on live camera feeds.

Real-time Object Detection Surveillance Software thumbnail
  • Fine-tuned and added GUI to an object detection neural network.
  • Increased frame throughput by 60% using multi-threading and multi-processing.
  • Deployed the software across six army bases for user testing.
Computer Vision Object Detection PyTorch PyQt UI/UX Multi-Processing
View More

Shoe Detection, Segmentation, and Stitching

A pipeline that overlays retailer shoes on top of user photos.

Shoe Detection, Segmentation, and Stitching thumbnail
  • Used YOLOv3 for shoe detection.
  • Used GrabCut with SIFT features convex hull for segmentation.
  • Stitched retailer shoes on top of user images by applying similarity transformations based on shoe contours.
Computer Vision Object Detection Object Segmentation Image Stitching Algorithm
View More GitHub

LSTM-based Classifier for Counter-Battery Radars

A neural network for quick projectile classification in counter-battery radars.

LSTM-based Classifier for Counter-Battery Radars thumbnail
  • Achieved 94.3% accuracy and 0.85 F1-score on real-world data collected by counter-battery radars.
  • Patent granted (10-2641022) in Korea as property of the Republic of Korea Army.
Time Series Classification PyTorch Patent-Granted
View More

Malicious Web Request Classification using NLP

  • Implemented and trained a custom four-layer BERT model on 45k web application firewall log entries, classifying them into one of nine classes of malicious activities.
  • Achieved 85.6% accuracy and 0.90 macro F1-score by using a random forest classifier on top of BERT embeddings.
Natural Language Processing Classification PyTorch

Multi-Vehicle Routing Algorithm [2nd Place]

A heuristic algorithm for computing simultaneous package delivery routes for multiple vehicles.

Multi-Vehicle Routing Algorithm [2nd Place] thumbnail
  • Won 2nd place (out of 103 teams) in the 2022 CJ Logistics Future Technology Challenge.
  • Achieved the maximum possible delivery count on the competition dataset (37 vehicles, 1,805 packages).
  • Decreased runtime by 20% using multi-threading, while maintaining deterministic results.
Algorithm C++ Multi-Threading
View More