The problem Public anomaly detector solves

There are innumerable cameras present in any public setting for facilitating security and it is an increasingly difficult task for humans to keep track of the same and detect anomalies taking place. This task can easily be delegated to Computer Vision models to notify security personnel about matters of interest in CCTV footage. Our project checks for usual activities in CCTV footage by employing a computer vision model trained on UCSD CCTV data. This can be used across the board to facilitate public security in various avenues. The project consists of a Tensorflow model based on CNNs, LSTMs and autoencoders and is integrated with a Flask backend and a Next JS Frontend to create a cohesive web platform.

Challenges we ran into

Our model was difficult to train due to the limited amount of computation power that personal systems offer hence limiting accuracy. As the backend receives video footage and the model takes input in image sequences, it is fairly difficult to achieve parity between the two, without running out of available system memory on a personal testing system thus we had to limit the number of frames in the test image sequence to 200. The backend is written in python on the Flask module to have easy integration with the model while the front end is written in Javascript on the NextJS framework which made it difficult to integrate the two.