Abstract
This work presents a novel and unique traffic
surveillance dataset, the MTID. When capturing data for traffic
surveillance, the positioning of the camera is crucial, and
depending on the task different approaches provide different
advantages. Multiple viewpoints, however, are rarely compared.
Our dataset gives the ability to analyze the difference between
two viewpoints in great detail.
A complex traffic scene has been captured simultaneously
from two different viewpoints: that of a camera mounted on
existing infrastructure, and an that of a drone. The frames
from each video capture have been synchronized in time and
all road users have been carefully annotated down to pixellevel accuracy. The dataset consists of 3100 frames from each
viewpoint, containing 18883 individual annotations on the pole
viewpoint, and 50274 individual annotations on the drone
viewpoint. The dataset is freely available online*.
Apart from the dataset, which is our main contribution, we
also provide benchmark detection results for four different
groups of road users for other researchers to compare their
results with. We show that the detection problem is challenging,
as we achieve mAPs of only 22.62% and 27.75% using a pretrained state-of-the-art detector on the two viewpoints.
surveillance dataset, the MTID. When capturing data for traffic
surveillance, the positioning of the camera is crucial, and
depending on the task different approaches provide different
advantages. Multiple viewpoints, however, are rarely compared.
Our dataset gives the ability to analyze the difference between
two viewpoints in great detail.
A complex traffic scene has been captured simultaneously
from two different viewpoints: that of a camera mounted on
existing infrastructure, and an that of a drone. The frames
from each video capture have been synchronized in time and
all road users have been carefully annotated down to pixellevel accuracy. The dataset consists of 3100 frames from each
viewpoint, containing 18883 individual annotations on the pole
viewpoint, and 50274 individual annotations on the drone
viewpoint. The dataset is freely available online*.
Apart from the dataset, which is our main contribution, we
also provide benchmark detection results for four different
groups of road users for other researchers to compare their
results with. We show that the detection problem is challenging,
as we achieve mAPs of only 22.62% and 27.75% using a pretrained state-of-the-art detector on the two viewpoints.
| Original language | English |
|---|---|
| Title of host publication | IEEE 23rd International Conference on Intelligent Transportation Systems 2020 |
| Number of pages | 6 |
| Publisher | IEEE |
| Publication date | 2020 |
| Article number | 9294694 |
| ISBN (Print) | 978-1-7281-4150-3 |
| ISBN (Electronic) | 978-1-7281-4149-7 |
| DOIs | |
| Publication status | Published - 2020 |
| Event | 23rd IEEE International Conference on Intelligent Transportation Systems 2020 - Duration: 20 Sept 2020 → 23 Sept 2020 https://ieee-itsc2020.org/ |
Conference
| Conference | 23rd IEEE International Conference on Intelligent Transportation Systems 2020 |
|---|---|
| Period | 20/09/2020 → 23/09/2020 |
| Internet address |
Fingerprint
Dive into the research topics of 'Presenting the Multi-View Traffic Intersection Dataset (MTID): A Detailed Traffic-Surveillance Dataset'. Together they form a unique fingerprint.Datasets
-
Multi-view Traffic Intersection Dataset (MTID)
Jensen, M. B. (Creator), Møgelmose, A. (Editor) & Moeslund, T. B. (Editor), Kaggle, 2020
https://vap.aau.dk/mtid/ and one more link, https://www.kaggle.com/datasets/andreasmoegelmose/multiview-traffic-intersection-dataset (show fewer)
Dataset