Thanks for your interesting in the project, in case of questions please mail us: contact@violencedetector.org
This is a real time violence detector project which will leverage the development of a family of new products for smart security devices. Today this technology is in the early stages of industrialization, therefore only big organizations have resources to develop it, that's why I've decided to opensource it. We can change it together, since this is too relevant to be restricted. The framework is written in python with openCV, the intention is keeping it as much multi plattform as possible.
It is:
- A real time violence detector framework built using computer vision, it comprehends but is not limited to:
- Dataset and it's composition,
- Image preprocessing,
- Network architecture: object detection / image classification,
- Keypoint extraction,
- Decision rules for the detection,
- Software must be hardware independent as much as much as possible
It is not:
- Final product development,
- Hardware related project,
- Privacy related discussion,
- It did not supposed to cover:
- Software not related to the detection,
- OS related matters as drivers, packages or external dependencies
Develop an opensource framework using computer vision to detect violence in real time. This real time violence detector framework will leverage the development of a family of new products for smart security devices.
General surveillance, anti terrorism, crowd management, assault detection, goods transportation protection...
Potential connection to police network for automatic crime detection and prevention.
The framework will take care about processing the image frame(s) and detect the existence of violence. With a simple output yes or no.
It is composed by python detection class file and a frozen model file, the target is being as much as possible platform invariant.
Framework intend to run in 2 modes:
- A) Framework runing at the same image acquisition device
- B) In a remote server via API call, sending images and returnind detection result
To keep it simple at the first moment the focus will be in the A mode, since the B mode should(at least this is the expectation) to re-utilize the majority of the code.
- Assault,
- Fight,
- Lynching,
- Hand gun,
- Knife,
- Terrorism,
- General violent behavior
If you want to contribute with the project, please take a look in the CONTRIBUTORS.md where you can find relevant additional information about the project.
In case of questions, please contact us contact@violencedetector.org
To run the method in the article(https://www.igi-global.com/article/simple-approach-for-violence-detection-in-real-time-videos-using-pose-estimation-with-azimuthal-displacement-and-centroid-distance-as-features/304462) you firstly need to install the OpenPose framework.
There are multiple installation options, this is the official Repo: https://github.com/CMU-Perceptual-Computing-Lab/openpose/blob/master/doc/installation/0_index.md
For this demo, I've used one child repo based on Tensorflow implementation(https://github.com/ZheC/tf-pose-estimation) which I've cloned using:
git clone https://www.github.com/ildoonet/tf-openpose
cd tf-openpose
pip3 install -r requirements.txt
Having the Openpose installed, I've created one additional method under the main class TfPoseEstimator to support the feature extraction per individual. You can replace the original file in the location(Pose_Estimation/tf-pose-estimation/tf_pose) by the enhanced one(estimator.py) available in this repo.
Before running the test(article.py) you need 3 steps:
- 1-Copy the frames(*) to be processed, I've added a test sample containing violence and non-violence scenes inside the repo image folder.
- 2-Copy the trained model(best_acc_final.keras) to the tf-pose-estimation folder.
- 3-Copy the python script article.py Now you can run the test passing the frames location as a parameter like that:
python3 article.py --file cam22.mp4*
Note: For keeping this demo simple, the code is limited to dealing with 2 individuals, nevertheless there is no limitation for this method be used with more individuals if the logic in article.py is enhanced for that.
The program will identify the individuals thru the 64 frames and present the last one with the inferred key points, after clicking ESC the model will receive the features extracted from the frames and you will see the array of the inference result per individual for the whole timeseries.
(*) To demonstrate the model generalization power, the images extracted for the demo are not from the same dataset used for training the model (M. Bianculli, N. Falcionelli, P. Sernani, S. Tomassini, P. Contardo, M. Lombardi, A.F. Dragoni, A dataset for automatic violence detection in videos, Data in Brief 33 (2020). doi:10.1016/j.dib.2020.106587).