Track vehicles and pedestrians without wiring YOLOv5 to DeepSort

What it does

This project is a Python integration layer that runs YOLOv5 object detection and DeepSort identity tracking together. It exposes a Detector class that ingests video frames and returns annotated results, specifically targeting vehicles and pedestrians. The goal is to let developers embed tracking and counting logic into their own applications without handling the two models’ integration manually.

The interesting bit

The value is in the encapsulation rather than novelty. The author hardwired the detection filter to person, car, and truck, and wrapped both the YOLOv5 preprocessing pipeline and DeepSort’s update cycle behind a single feedCap() method. It is essentially a ready-made sandwich of two well-known CV components, which is useful if you need exactly that filling.

Key highlights

• Bundles YOLOv5 inference and DeepSort tracking behind one Detector class.
• API accepts a raw BGR frame and returns a dict containing the annotated frame via feedCap().
• Supports swapping in custom YOLOv5 weights if you train your own detector.
• Automatically falls back to CPU if CUDA is unavailable.
• Author now maintains a YOLOv11 successor, though this YOLOv5 branch remains available.

Caveats

• The shown detect() method hardcodes class labels to person, car, and truck, so tracking other categories requires editing the source.
• The feedCap() interface includes an opaque func_status argument whose purpose is not documented beyond a headpose placeholder set to None.
• The project is effectively in maintenance mode; the README prominently links to a newer YOLOv11-DeepSort repository.

Verdict

Worth a look if you need a quick, drop-in vehicle or pedestrian tracker for an existing Python codebase and are comfortable reading source to extend it. Skip it if you want a generic multi-class tracker or an actively maintained framework.