README.md

[CVPRW2023] "Leveraging Future Trajectory Prediction for Multi-Camera People Tracking"

Track1: Multi-Camera People Tracking

The official repository for 7th NVIDIA AI City Challenge

Pipeline Overview

Environment

We run on 2 NVIDIA A6000 GPUs.

• Linux or macOS
• Python 3.7+ (Python 3.8 in our envs)
• PyTorch 1.9+ (1.11.0 in our envs)
• CUDA 10.2+ (CUDA 11.3 in our envs)
• mmcv-full==1.7.1 (MMCV)

Installation

• Step #1. Create environment (recommend environment)

conda env create --file environment.yaml
conda activate scit

• Step #2. Install packages

sh setup.sh

Train

Object Detection

• Train on Dataset

  We use mmdetection's YOLOX-X model.

  Follow mmdetection's guideline for training.

• Pretrained

  Pretrained person detection model weights on NVIDIA Omniverse Dataset from 2023 AI City Challenge Track1.

  pretrained-YOLOX-X.pth

Keypoint Detection

• Pretrained

  We directly use yolov7-pose-estimation's pretrained pose estimation model.

  You can download pretrained pose estimation model from their git page.

Trajectory Prediction

• Train on Dataset

  We use Social-Implicit model.

• Pretrained

  Pretrained trajectory prediction model weights on NVIDIA Omniverse Dataset from 2023 AI City Challenge Track1.

  pretrained-social-implicit.pth

Demo

• Step #1. Single-Camera Tracking.

sh run_scmt.sh

• Step #2. Multi-Camera Tracking (Association). Here is the homography_list.pkl.

sh run_mcmt.sh

Citation

@InProceedings{Jeon_2023_CVPR,
    author    = {Jeon, Yuntae and Tran, Dai Quoc and Park, Minsoo and Park, Seunghee},
    title     = {Leveraging Future Trajectory Prediction for Multi-Camera People Tracking},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops},
    month     = {June},
    year      = {2023},
    pages     = {5398-5407}
}