Project 05

Teknofest Traffic Computer Vision

A classical image-processing pipeline for traffic scenes — 698 labeled images processed with OpenCV, no deep learning, advanced to the competition's second stage.

§ 01 Problem

Detect and preprocess traffic elements under genuinely hostile conditions — variable light, motion blur, cluttered scenes — where the naive move is to throw a pretrained network at it and the constraint was to understand the image formation instead.

§ 02 Approach

  • Built a modular classical pipeline where each stage is a separate, swappable file: thresholding (src/threshold.py), morphology (morph.py), noise removal (noise.py), blending (blend.py) and filtering (filter.py) — so failures could be localized to a stage instead of debugging a monolith.
  • Ran the full 698-image dataset through the pipeline; its cleaned output became the input for the team's second-stage submission.

§ 03 Outcome

Advanced to stage two of the competition. The lasting value was lower-level: learning what thresholding and morphology actually do to an image before ever letting a network hide it.

§ 04 Evidence

evidenceteknofest-traffic-cv — src/, 698-image dataset
Traffic scene, pipeline sample
fig. 1 — Traffic scene, pipeline sample
Traffic scene, pipeline sample
fig. 2 — Traffic scene, pipeline sample
Traffic scene, pipeline sample
fig. 3 — Traffic scene, pipeline sample
Traffic scene, pipeline sample
fig. 4 — Traffic scene, pipeline sample