Furthering the field of CV.
What it does
ATLAS is a multi-platform, automated application that caters to global users, allowing them to upload videos from the cloud and integrate their social media to then search their videos for specific objects using keywords and phrases. ATLAS then displays the videos containing the object that they searched for. ATLAS leverages the latest computer vision research to identify objects that provide semantic phrases and words that categorize objects in videos. For example, a user could sync videos from their Facebook, Instagram, and Google Drive and then search for “bicycle”. Every video with a bicycle from different sources will then be displayed to the user. This is also applicable on larger scale. For instance, on January 27th, just outside the Johns Hopkins’ campus, there was an attempted kidnapping on a female student. A man in a minivan approached her with a gun, telling her to get in the car. She didn’t listen and kept walking, later reporting it to campus security. It took at least two days, January 29th, to process the video feeds from traffic and security cameras to collect more information to help identify the man. However, if security had used ATLAS, they potentially could have identified the minivan much quicker by searching “van”. Expediting the Identification the van sooner could have lead to actually finding the man responsible for the attempted kidnapping.
How we built it
We used TensorFlow libraries and other Python libraries to configure a strong CV analysis program that provides object recognition and labeling, to which we also interface with.
Challenges we ran into
Complete integration of the front end communication with back end sever application. TensorFlow is also an arduous library to implement and comes with many advanced features that did cloud our original thoughts. Writing a shell script to intermittently return frames of a video and then implement weighted graph-based data structure used for duplicate analysis and hashing.
Accomplishments that we're proud of
We integrated an entire backend CV system that allows to do advanced image recognition and recall using JSON tags.
What we learned
We learned to communicate and think together flexibly, which directly corresponded to how our project worked technically.
What's next for ATLAS
We would like to find a way for ATLAS to recognize verbs. By aggregating all the objects, we hope to relate objects to form action phrases. We would also like to use parallel programming, to harness GPU’s computational power, to reduce processing time. In addition, we would like to target to a more specific audience and perhaps use ATLAS in government surveillance programs and medical services.