IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW) · 2023
Leveraging Future Trajectory Prediction for Multi-Camera People Tracking
Yuntae Jeon, Dai Quoc Tran, Minsoo Park, Seunghee Park
A real-time, low-computation multi-camera people tracking system that predicts future trajectories to match people across cameras, without appearance-based re-identification (ReID)*.
* ReID (re-identification) matches identities by comparing visual features extracted from detected bounding boxes.
Challenges and Solutions
Track 1 of the AI City Challenge asks for detecting and tracking people across multiple camera views, known as Multi-Camera Multi-Target (MCMT) tracking.
Challenges
- Kalman filter-based methods are weak in object overlapping.
- ReID-based methods are weak when the appearance information of objects is not clear.
- ReID-based methods are not proper for real-time application.
Our Solutions
- Use trajectory prediction for Multi-Camera Multi-Target (MCMT) tracking.
- Utilize only the motion information from tracklets of each individual camera for MCMT tracking.
- Utilize lightweight CNN-based trajectory prediction without ReID for real-world & real-time application of MCMT tracking.
Abstract
Artificial intelligence-based surveillance system, one of the essential systems for smart cities, plays a critical role in ensuring the safety and well-being of individuals. In this paper, we propose a real-time, low-computation cost Multi-Camera Multi-Target (MCMT) tracking system for people, leveraging deep-learning-based trajectory prediction with spatial-temporal information and social information.
By predicting people's future trajectories, our algorithm effectively handles object occlusion problems and maintains accurate tracking while keeping computational costs low. Our approach addresses object occlusion without relying on computationally expensive re-identification, and improves MCMT tracking performance using graph-based tracklet representation, and spectral clustering.
As a result, our proposed approach is tested on the 2023 AI City Challenge Track 1 test dataset, automatically generated on the NVIDIA Omniverse Platform, our method achieves an IDF1 score of 0.6171 and real-time performance at 27.6 FPS. Code and pre-trained models are publicly available at https://github.com/yuntaeJ/SCIT-MCMT-Tracking.
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 = {5399-5408},
doi = {10.1109/CVPRW59228.2023.00570}
}