ramiyashree/deep-see-crime

An AI based surveillance system to track down suspects using gestures and facial attributes using UCF Crime Dataset 128 hours long real-world surveillance videos 13 realistic anomalies includes fighting, assault, road accidents.

Smart India Hackathon 2020 MS335 CODE MONK

DEEP SEE CRIME

Special Thanks

Thanks Swarna Swapna for mentoring and supporting in Deep Learning model creation.

The Goal

Can state-of-the-art deep neural networks “See” violence in images and videos ?? To signal an activity that deviates normal patterns with time window. Video annotation, Video retrieval, and Real-time monitoring. Identify and track down the suspects. Note: Real-world anomalous events are complicated and diverse. It is difficult to classify all of the possible anomalous events.

The Data

UCF Crime Dataset 128 hours long real-world surveillance videos 13 realistic anomalies includes fighting, assault, road accidents. Weakly labelled training videos. i.e. data is labelled video level, but which duration isn’t tagged. https://www.crcv.ucf.edu/projects/real-world/

The Method

Our approach considers anomalous and normal events for improper behaviour detection.

1. Formulates anomaly score for a video clip and provides time window of the crime event.
2. Classifies crimes based on Action Recognition task for Video Retrieval and monitoring.
3. Tag the suspects present in the time frame and track them.

Overall Pipeline

Technology Stack

Pipeline Backend

Python 3, Open CV, Tensorflow, TF Records, Pytorch, SqLite Database

Deep Learning Models

Module 1 - Inflated 3D CNN Model, Module 2 - PySlowFast Model, Module 3 - OpenPose + DeepSort

Web Front and Backend

Java Script, HTML, CSS, Bootstrap, Django Restful API, Docker Containers