Directional headcounts from a webcam and some glue code

What it does This project detects and tracks six default classes—pedestrians, bicycles, cars, motorcycles, buses, and trucks—then tallies how many cross a virtual boundary in each direction. It is essentially a practical integration of existing detection and tracking code, built as a ready-to-run demo rather than a general-purpose framework.

The interesting bit Instead of just drawing bounding boxes, it uses two polygons in main.py as directional gates, giving separate “in” and “out” counts. That turns the standard YOLOv5-plus-DeepSORT stack into a simple traffic analytics tool, though you tune it by editing source coordinates rather than waving a config file at it.

Key highlights

• Detects and tracks six default classes using YOLOv5 and DeepSORT, with categories adjustable in detector.py
• Splits traffic into “in” and “out” counts using two polygons defined in main.py
• Ships with a Bilibili video demo and a 140 MB test video (hosted on Baidu Pan)
• Acts as a use-case-specific wrapper around three existing repositories

Caveats

• All configuration shown in the README requires editing specific lines in main.py and detector.py; no command-line interface or config file is mentioned
• The README and code comments are in Chinese
• The repository contains weights and video files, so the maintainers warn that cloning may be slow

Verdict Worth a look if you want a hackable, offline baseline for directional people or vehicle counting. Skip it if you need a production-grade, configurable video analytics platform.