NEWS
- Our paper has been accepted by IEEE Access in Feb, 2025! The download link is now available.
- The project page is now open!
OVERVIEW
The CADG dataset is a new dataset for Crowd Anomaly detection from Drone and Ground views.
The dataset provides a sophisticated simulation of how people would behave during an incident or accident in a public space, allowing the evaluation of various crowd activity detection methods for application in real-world situations. The dataset is not only the largest and most comprehensive dataset currently available for crowd anomaly detection, but also provides videos captured by both a moving drone and fixed cameras.
From the analysis of crime statistics in public spaces and the perspective of importance in crowd monitoring, we have selected four important crowd anomaly activities, i.e., (i) fight, (ii) theft, (iii) illness, and (iv) dispersion, as shown in the following examples. All videos were recorded in a gymnasium with up to 12 actors.
The normal activities include a wide variety of daily activities such as random walking, standing, texting while walking, and talking with several people. For the anomaly activities, a few randomly selected people played particularly important roles, such as people suddenly falling ill or robbers. The rest of the actors intervened by, for example, walking around while paying attention, stopping to observe the incident, or talking to the afflicted or victims.
The dataset provides a sophisticated simulation of how people would behave during an incident or accident in a public space, allowing the evaluation of various crowd activity detection methods for application in real-world situations. The dataset is not only the largest and most comprehensive dataset currently available for crowd anomaly detection, but also provides videos captured by both a moving drone and fixed cameras.
From the analysis of crime statistics in public spaces and the perspective of importance in crowd monitoring, we have selected four important crowd anomaly activities, i.e., (i) fight, (ii) theft, (iii) illness, and (iv) dispersion, as shown in the following examples. All videos were recorded in a gymnasium with up to 12 actors.
The normal activities include a wide variety of daily activities such as random walking, standing, texting while walking, and talking with several people. For the anomaly activities, a few randomly selected people played particularly important roles, such as people suddenly falling ill or robbers. The rest of the actors intervened by, for example, walking around while paying attention, stopping to observe the incident, or talking to the afflicted or victims.
SAMPLES
* The videos shown above have been trimmed from the original clips.
DOWNLOAD
If you wish to access the dataset, please read and understand the following Terms of Use, and send an email to cadg.dataset [at] gmail.com. By sending the email, you are deemed to have agreed to the Terms of Use. We will then send you the link to download the dataset!
Terms of Use
- The use of this dataset is strictly limited to academic research ONLY. Any direct or indirect commercial use is prohibited.
- To protect privacy, all videos in this dataset are permitted to be used ONLY for demonstrations in academic publications and presentations.
- Any reproduction, modification, combination with other data, or distribution of this dataset, whether partially or in its entirety, including any models constructed using this dataset, is strictly prohibited without the express permission of the data publisher. Any such unauthorized use will be considered illegal.
- Hitachi, Ltd. assumes no responsibility for any losses, expenses, or damages caused by the use of this dataset. All users agree to indemnify, defend, and hold harmless the data collection participants, Hitachi, Ltd., and its officers, employees, and agents, both individually and collectively.
- Any usage that violates the above terms will result in the immediate termination of the usage license. Users must permanently delete all content related to this dataset upon termination.
Citation
If you use this dataset in your research, please cite it as follows:
@article{ito2025crowd,
title = {Crowd Anomaly Detection from Drone and Ground},
author = {Ito, Yoshiki and Sasaki, Tatsuya and Kondo, Satoshi},
journal = {IEEE Access},
volume = {13},
pages = {28824--28835},
year = {2025},
publisher = {IEEE}
}