The project of Anti-UAV is released under the MIT License.
✅ 07 June 2024: We have released the Jittor version of Anti-UAV for domestic hardware support and inference speed up. 😊
✅ 04 July 2024: We have released the Jittor version of Anti-UAV410 for domestic hardware support and inference speed up. 😊
To our best knowledge, we are the first to propose a new Anti-UAV task, corresponding datasets, evaluation metrics and baseline methods, with both PyTorch and Jittor support.
- Task Definition
- Motivation
- Environment Setup
- Data Preparation
- Evaluation Metrics
- Training and Inference
- Notebook
- Model Zoo
- FAQs
- Achievement
- Citation
Anti-UAV refers to discovering, detecting, recognizing, and tracking Unmanned Aerial Vehicle (UAV) targets in the wild and simultaneously estimate the tracking states of the targets given RGB and/or Thermal Infrared (IR) videos. When the target disappears, an invisible mark of the target needs to be given. A lot of higher-level applications can be founded upon anti-UAV, such as security of important area, defense against UAV attack, and automated continuous protection from potential threat caused by UAV instrusion.
-
The Anti-UAV project of Institute of North Electronic Equipment, Beijing, China is proposed to push the frontiers of discovery, detection and tracking of UAVs in the wild.
-
Recently, UAV is growing rapidly in a wide range of consumer communications and networks with their autonomy, flexibility, and a broad range of application domains. UAV applications offer possible civil and public domain applications in which single or multiple UAVs may be used. At the same time, we also need to be aware of the potential threat to airspace safty caused by UAV intrusion. Earlier this year, multiple instances of drone sightings halted air traffic at airports, leading to significant economic losses for airlines.
-
Currently, in computer vision community, there is no high-quality benchmark for anti-UAV captured in real-world dynamic scenarios. To mitigate this gap, this project presents a new dataset, evaluation metric, and baseline method for the area of discovering, detecting, recognizing, and tracking UAVs. The dataset consists of high quality, Full HD video sequences (both RGB and IR), spanning multiple occurrences of multi-scale UAVs, densely annotated with bounding boxes, attributes, and flags indicating whether the target exists or not in each frame.
Refer to: 3rd_Anti-UAV_CVPR2023.
Note:
- We recommend that you use
python == 3.8, or conflicts may occur; - For those using NVIDIA RTX 30 & 40 series GPUs, the cuda version 11.8 has been tested successfully;
- When you run the command
pip install -r requirements/cv.txt, there might be some problems, like unable to find bmt_clipit, clip, panopticapi, videofeatures_clipit. And these errors can be ignored; - You need to install jibjpeg4py library and
jittor == 1.3.8.5.
Currently, we offer three public datasets for the ANTI-UAV task.
- Folder Tree Diagram
- Scenario Variations: Compared to the previous challenge, we further enlarge the dataset this year by adding more challenging video sequences with dynamic backgrounds, complex movements, and tiny-scale targets, such that the resulting new dataset covers a greater variety of scenarios with multi-scale UAVs. Examples are shown as follows.
- Download:
| Dataset | Google Drive | Baidu Drive |
|---|---|---|
| Anti-UAV300 | link | Password:sagx |
| Anti-UAV410 | N/A | Password:wfds |
| Anti-UAV600 | modelscope | N/A |
- Please note that 410 and 600 versions only contain IR videos while 300 version contains both RGB videos and IR videos. In this release, the model is capable of dealing with both RGB data and IR data, so we recommend that you use the 300 version dataset.
- Please refer to our Anti-UAV v1 paper and Anti-UAV v3 paper for more details (WeChat News).
We define the tracking accuracy as:
For frame t, IoU_t is Intersection over Union (IoU) between the predicted tracking box and its corresponding ground-truth box, p_t is the predicted visibility flag, it equals 1 when the predicted box is empty and 0 otherwise. The v_t is the ground-truth visibility flag of the target, the indicator function δ(v_t>0) equals 1 when v_t > 0 and 0 otherwise. The accuracy is averaged over all frames in a sequence, T indicates total frames and T^* denotes the number of frames corresponding to the presence of the target in the ground-truth.
- Training
Currently, we have some problems with training in Jittor, but you can still try to run the commands as follows:
cd {PROJECT_ROOT}/anti_uav_jittor
python ltr/run_training.py modal modalOr you can train the model with PyTorch.
Also, if you have any suggestions on how to do it, feel free to open an issue!
- Inference
In the root path of the project, run the command python pysot_toolkit/test.py
We provide a demo notebook in anti_uav_jittor/demo.ipynb that can help developers better understand the workflow of this demo.
🐒
Keep updating...
We will keep updating this section, so feel free to open an issue.
1. Q: When I run the training command, there is an error indicating that setting an array element with a sequence, the requested array has an inhomogeneous shape.
A: In AntiFusion.py, set `visible_data = np.array, dtype = object`.- CVPR 2020 Anti-UAV Workshop & Challenge
- ICCV 2021 Anti-UAV Workshop & Challenge
- CVPR 2023 Anti-UAV Workshop & Challenge
-
Please consult and consider citing the following papers:
@article{zhang2023review, title={Review and Analysis of RGBT Single Object Tracking Methods: A Fusion Perspective}, author={Zhang, ZhiHao and Wang, Jun and Zang, Zhuli and Jin, Lei and Li, Shengjie and Wu, Hao and Zhao, Jian and Bo, Zhang}, journal={T-OMM}, year={2023} } @article{huang2023anti, title={Anti-UAV410: A Thermal Infrared Benchmark and Customized Scheme for Tracking Drones in the Wild}, author={Huang, Bo and Li, Jianan and Chen, Junjie and Wang, Gang and Zhao, Jian and Xu, Tingfa}, journal={T-PAMI}, year={2023} } @inproceedings{zhang2023modality, title={Modality Meets Long-Term Tracker: A Siamese Dual Fusion Framework for Tracking UAV}, author={Zhang, Zhihao and Jin, Lei and Li, Shengjie and Xia, JianQiang and Wang, Jun and Li, Zun and Zhu, Zheng and Yang, Wenhan and Zhang, PengFei and Zhao, Jian and others}, booktitle={ICIP 2023}, year={2023} } @article{jiang2021anti, title={Anti-UAV: a large-scale benchmark for vision-based UAV tracking}, author={Jiang, Nan and Wang, Kuiran and Peng, Xiaoke and Yu, Xuehui and Wang, Qiang and Xing, Junliang and Li, Guorong and Ye, Qixiang and Jiao, Jianbin and Han, Zhenjun and others}, journal={T-MM}, year={2021} } @article{zhao20212nd, title={The 2nd anti-UAV workshop \& challenge: methods and results}, author={Zhao, Jian and Wang, Gang and Li, Jianan and Jin, Lei and Fan, Nana and Wang, Min and Wang, Xiaojuan and Yong, Ting and Deng, Yafeng and Guo, Yandong and others}, journal={arXiv preprint arXiv:2108.09909}, year={2021} } @article{zhu2023evidential, title={Evidential Detection and Tracking Collaboration: New Problem, Benchmark and Algorithm for Robust Anti-UAV System}, author={Zhu, Xue-Feng and Xu, Tianyang and Zhao, Jian and Liu, Jia-Wei and Wang, Kai and Wang, Gang and Li, Jianan and Zhang, Zhihao and Wang, Qiang and Jin, Lei and others}, journal={arXiv preprint arXiv:2306.15767}, year={2023} } @article{zhao20233rd, title={The 3rd Anti-UAV Workshop \& Challenge: Methods and Results}, author={Zhao, Jian and Li, Jianan and Jin, Lei and Chu, Jiaming and Zhang, Zhihao and Wang, Jun and Xia, Jiangqiang and Wang, Kai and Liu, Yang and Gulshad, Sadaf and others}, journal={arXiv preprint arXiv:2305.07290}, year={2023} } @article{zhao20233rd, title={The 3rd Anti-UAV Workshop \& Challenge: Methods and Results}, author={Zhao, Jian and Li, Jianan and Jin, Lei and Chu, Jiaming and Zhang, Zhihao and Wang, Jun and Xia, Jiangqiang and Wang, Kai and Liu, Yang and Gulshad, Sadaf and others}, journal={arXiv preprint arXiv:2305.07290}, year={2023} }