Sample Frame 1
Sample Frame 2
Video surveillance feeds run 24-7. The workers who monitor these feeds to pull salient information take hours on end to review this video, even at several times playback speed, often at greatly reduced accuracy. Our team knew there should be an intelligent way of analyzing these videos.
What it does
QuickLook automatically analyzes these video feeds, identifying salient points in the time series and flagging periods that may be of significance to investigators.
How we built it
We used open data video data sets such as Virat and Canadian Open Data. Our team took Faster RCNN Tensorflow models to recognize objects in our video data sets, and added heuristics to gather salient features from the analyzed videos.
Challenges we ran into
Scalability: Analyzing every frame in vast volumes of video data proved to be difficult. We bypassed this by utilizing a cluster of GPUs and parallelizing the work loads.
Processing and insight extraction: Processing the raw data to gather insights and recognize patterns/discrepancies within the video feeds.
Accomplishments that we're proud of
Data adaptation: Data gathering was made agnostic to the video and to the final purpose. The software we've created can be applied and (and further optimized) to perform in other settings, such as: convenience stores and school surveillance feeds.
Excellent UI/UX Design: The team created an end product with an interface designed for ease of use. By combining a simple UI with select important elements, we ensure that the 'assistive' aspect of our product delivers.
First-of-its-kind approach: QuickLook is a first-of-its-kind design that implements deep learning and Faster RCNN into activity and object detection for security feeds. By taking this approach, we hope to make progress in identifying key events, and further developing key strategies to optimize our processes (and even predict event occurences).
What's next for QuickLook
Our team is looking into deeper levels of object and activity recognition accuracy, and implementing a predictive event model to compliment our currrent video analytics product.