mcv-c6-2025-team8

Multi-Camera Vehicle Tracking

Overview

This project focuses on multi-camera vehicle tracking using object detection, tracking, and re-identification (ReID). The pipeline involves:

• Object Detection using YOLO.

• Tracking using SORT.

• Cross-camera Re-identification (ReID) to match objects across multiple camera views.

• Filtering of detections based on matched object pairs.

• Evaluation using HOTA and IDF1 metrics.

Folder Structure

project_root/
│── output/
│   ├── detection/
│   │   ├── S01/
│   │   │   ├── detections_1.txt
│   │   │   ├── ...
│   │   ├── S03/
│   │   │   ├── detections_10.txt
│   │   │   ├── ...
│   │   ├── S04/
│   │   │   ├── detections_16.txt
│   │   │   ├── ...
│   ├── filtered_detection/
│   │   ├── S01/
│   │   │   ├── filtered_detections_1.txt
│   │   │   ├── ...
│   │   ├── S03/
│   │   │   ├── filtered_detections_10.txt
│   │   │   ├── ...
│   │   ├── S04/
│   │   │   ├── filtered_detections_16.txt
│   │   │   ├── ...
│── matched_objects_S01.txt
│── matched_objects_S03.txt
│── matched_objects_S04.txt
│── filtering.py
│── matching.py
│── HOTA-IDF1.py
│── README.md

Filtering Detections

The filtering.py script filters detections based on object matching.

Usage:

python filtering.py

This script:

1. Loads the matched object pairs from matched_objects_{sequence}.txt.
2. Filters detections from output/detection/{sequence}/detections_{camera_id}.txt.
3. Saves filtered detections to output/filtered_detection/{sequence}/filtered_detections_{camera_id}.txt.

Evaluation with HOTA and IDF1

The HOTA-IDF1.py script evaluates the filtered detections.

Usage:

python HOTA-IDF1.py

This script computes:

• HOTA (Higher Order Tracking Accuracy): Evaluates both detection and association accuracy.
• IDF1 (ID F-score): Measures identity preservation in tracking.

Results

After filtering, the output filtered detections will be available in output/filtered_detection/. The evaluation metrics will be printed on the console.

Requirements

• Python 3.x
• Ultralytics YOLO
• torchreid
• collections
• torch
• scipy
• torchvision

Acknowledgements

This project uses data from the AI City Challenge and utilizes the tracking algorithm SORT. If you use this project or data, please cite the AI City Challenge.

• AI City Challenge 2022
  M. Naphade, J. Tang, S. Wang, L. Chang, A. Sharma, S. Birchfield, A. Das, R. Zhang, J. P. C. Valentin, A. Kirillov, D. Park, and Z. Tian,
  "AI City Challenge 2022", IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2022.
  AI City Challenge

• SORT (Simple Online and Realtime Tracker)
  A. Bewley, Z. Ge, L. Ott, F. Ramos, and B. Upcroft,
  "Simple Online and Realtime Tracker", IEEE International Conference on Image Processing (ICIP), 2016.
  [Paper] | [GitHub Repo]