Multi-Vehicle Tracking Using Heterogeneous Neural Networks for Appearance And Motion Features
- Authors
- Abdallah, Mohamed S.; Han, Da Sol; Kim, HyungWon
- Issue Date
- Dec-2022
- Publisher
- SPRINGER
- Keywords
- Convolution Neural network (CNN); KL divergence; Multiple Object Tracking (MOT); Object detection; Triplet loss; Vehicle tracking
- Citation
- INTERNATIONAL JOURNAL OF INTELLIGENT TRANSPORTATION SYSTEMS RESEARCH, v.20, no.3, pp.720 - 733
- Journal Title
- INTERNATIONAL JOURNAL OF INTELLIGENT TRANSPORTATION SYSTEMS RESEARCH
- Volume
- 20
- Number
- 3
- Start Page
- 720
- End Page
- 733
- URI
- https://scholarworks.bwise.kr/gachon/handle/2020.sw.gachon/86131
- DOI
- 10.1007/s13177-022-00320-6
- ISSN
- 1348-8503
- Abstract
- This paper presents a multi-vehicle tracking algorithm using appearance feature and motion history based on heterogeneous deep learning aimed at autonomous driving applications. Our proposed multi-vehicle tracking model follows the tracking-by-detection paradigm. To track multiple vehicles, we utilize the appearance and motion features of the target vehicles in consecutive frames. The proposed multi-vehicle tracking system employs a deep convolutional neural network, which is trained with a triplet loss minimization method to extract appearance features. The key contribution of the proposed method lies in a Long Short-Term Memory (LSTM) with a fully connected layer that accurately predicts the probability distribution of the next appearance and motion features of tracked objects. We constructed a multi-vehicle tracking dataset from various real road traffic using a camera sensor on a vehicle. To evaluate our proposed algorithm, we use several multi-target tracking datasets from the KITTI object tracking benchmark, which is a Public tracking dataset, as well as our evaluation dataset. Experimental results demonstrate that the proposed multi-vehicle tracking algorithm achieves a MOTA of 84.5% and MOTP 86.3% on the KITTI tracking dataset, and a MOTA of 81.8% and MOTP 84.8% on our evaluation dataset, an improvement of 8.6% and 9.6% over the previous methods.
- Files in This Item
- There are no files associated with this item.
- Appears in
Collections - ETC > 1. Journal Articles
![qrcode](https://api.qrserver.com/v1/create-qr-code/?size=55x55&data=https://scholarworks.bwise.kr/gachon/handle/2020.sw.gachon/86131)
Items in ScholarWorks are protected by copyright, with all rights reserved, unless otherwise indicated.