Wiring YOLOv4 into DeepSORT so you don’t have to

What it does

This project is integration glue: it marries YOLOv4 detections with DeepSORT to maintain persistent identity tracks across video frames. Hand it a video file or webcam stream and it returns an annotated video with bounding boxes and unique object IDs. It also swaps in YOLOv4-Tiny or YOLOv3 if you need faster inference.

The interesting bit

The heavy lifting is borrowed—weights are converted from Darknet to TensorFlow, and the tracker is vanilla DeepSORT—but the convenience is real. The authors wrapped detection, tracking, and video encoding into one pipeline, and even threw in a Google Colab notebook so you can skip local dependency wrangling.

Key highlights

• End-to-end TensorFlow pipeline from YOLOv4 detection to DeepSORT tracking
• Supports YOLOv4-Tiny for faster inference at a self-described accuracy cost
• Can restrict tracking to specific COCO classes by editing an allowed_classes list
• Outputs annotated video with tracked object IDs
• Provides a Google Colab notebook for cloud-based testing

Caveats

• Class filtering requires manually editing object_tracker.py rather than using a command-line flag
• The README pins CUDA 10.1, signaling the TensorFlow version is dated
• Pre-trained weights must be fetched separately from Google Drive

Verdict

A solid starting point for researchers or prototypers who need multi-object tracking on video without building the YOLO-to-DeepSORT bridge from scratch. Not for anyone looking for a maintained, production-ready tracking service.