Multi-Target Tracker for Low Light Vision

Introduction

This repository contains implementation for Multi-Target Tracker for Low Light Vision. LLV tracker is a multi-object tracker for TIR images with a focus on simple and real-time efficient algorithmic solution. We base our solution on DeepSORT algorithm and extend it to TIR tracking of both, pedestrians and vehicles. To adopt DeepSORT tracker, we design an appearance descriptor suitable for association problem of TIR images. Furthermore, to address the problem of missing association and detection, we propose a fusion block to merge short tracklets belonging to the same object in one track. We evaluate the tracker on CAMEL dataset and experimentally on the sequences we collected using an IR-camera.

Running LLV Tracker

python run_tracking.py \
    --sequence_dir (Path to sequence directory)\
    --min_confidence (Disregard detections lower threshold)\
    --device (cuda device, i.e. 0 or 0,1,2,3 or cpu)\
    --k (number of consequitive frames to count track as active)

Overview of Repository

The main entry point is in run_tracking.py. This file runs the tracker on a specified sequence.

Code Files:

• descriptor.py: Functions to compute descriptor.
• kalman_filter.py: A Kalman filter implementation and concrete parametrization for image space filtering.
• track.py: The track class contains single-target track data such as Kalman state, number of hits, misses, hit streak, feature matrix, etc.
• tracker.py: Multi-target tracker class.
• utils_functions.py: Helping functions.

Citing LLV Tracker

If you find this repo useful in your research, please consider citing the following paper:

@inproceedings{madjid2023multi,
  title={Multi-Target Tracker for Low Light Vision},
  author={Madjid, Nadya Abdel and Sharma, Arjun and Hassan, Bilal and Werghi, Naoufel and Dias, Jorge and Khonji, Majid},
  booktitle={2023 21st International Conference on Advanced Robotics (ICAR)},
  pages={252--257},
  year={2023},
  organization={IEEE}
}