-
Neutral Activity Recognition
-
Fighting Activity Recognition
-
Assault Activity Recognition
Inspiration
We were inspired by the millions of Mexicans affected by the crime rates in 2002- 2010 in Mexico, and also because we experienced life handed that kind of situation
What it does
We classify videos and rate by crimes, so we know when a store or a person is rob, that way we call authorities and ambulance beforehand. We also give the authorities information about the place, time, a video and a classify situation.
How we built it
We use python with TensorFlow and Keras, also we used the structure of ResNet50 with an extra hidden layer of 512 density, and an SGD optimizer. After that, we took an average of the probabilities between a certain range of frames using queues and starting classifying the videos. In the end, we pass down the information to a JSON for future scalability.
Challenges we ran into
Classifying videos without using an RNN or LSTM is difficult to concrete, but we manage to get good accuracy in the videos, also we challenge with finding the good loss rate for the network.
Accomplishments that we're proud of
We get a 70% accuracy in the assault class. We came with bare experience of Video Classification and CNN but managed to get a good project.
What we learned
We learn used more CNN tool´s and solve real problems with it.
What's next for Udyat
We want to add an LSTM or an RNN to increase the accuracy quality of the predictions, also to give our NN a sense of time and space, so it can classify using past frames an give it a context.
Log in or sign up for Devpost to join the conversation.