cornhack2019

• 

  Violence detection

Inspiration

• As a victim of bullying, I wish someone was there to stop it. The current problem with safety is that with a large student body on campus, it's impossible to help every individual who is in danger.

How we built it

• We started tackling the problem by looking at our University of Nebraska students community.
• Models were trained on a UCF Fighting dataset. We acquired images data of roughly 3000 violence images, which was then filtered into violence and non-violence. We use Tensorflow as the backend to handle video frame predicting. The model is stacked with 4 interlacing layers of convolutional, Batchnorm and dropout. A front end framework using JS and HTML to deploy our deep learning application. We simulate live video feeds from security cameras by having a pre-labeled video by our model embedded into difference region of the map.

What it does?

• Our projects integrate with a camera system process video feeds frames by frame and predict if a violence scene happen. When a violence scene happens a notification is sent.
• However, using deep learning technology, we built a system that can detect violence as soon as it happens. This could integrate with surveillance security cameras to identify at-risk situations quickly and can be used to notify important persons to help

Challenges we ran into

• We had an issue training our deep-learning model. With our limitation of hardware, The validation process of hyperparameter tuning and training process consumes a large amount of times. Lacks of images data cause our model to fluctuate with an accuracy of 70% on trainset.
• Many of us do not have any background in javascript and front end design. The integration between the backend process and front-end is our biggest pull-back.

Accomplishments that we're proud of

• With a limited amount of times(24hrs) and limited resources, we were able to hacks and tackle a real-life problem.

What we learned

• As software developers, we learned how to collaborate in a limited time environment. With a lack of resources, we learned to work around and find an alternative solution.

What's next for cornhack2019

• For the future, we would like to improve our model by augmenting more features like detecting medical emergencies,...
• Many security systems use drones to manage a large body of communities. Our system can help integrate risk situations quickly and can be used to notify important personnel to help

Built With

• css
• html
• javascript
• jupyter-notebook
• python
• tensorflow