Summarize hours of footage shot by static CCTV cameras into a short clip that shows all events as if they're occurring concurrently with timestamps as shown in the demo.
In the modern times of security, Surveillance cameras are making their way into the domestic markets as well. It becomes a very tedious task to go through tens of hours of footage to check for any suspicious activity in the area. It also takes a lot of memory to store these long videos, and sometimes the resolution of the video is compromised to save memory. There is a need for surveillance solutions that can abridge such long footage into short videos while still retaining important events.
CCTV-Recap can summarize hours of footage shot by static CCTV cameras into a short clip that shows all interesting events simultaneously. The program identifies and tracks all moving objects present in the video and overlays these events on a single clip, while also showing the timestamps for each event, thus letting the user perform surveillance for multiple events.
-
We use a moving average of all frames in the video to approximate the background. As disturbances don't often last throughout the duration of the video, only the background's pixel values are reflected in the final average.
-
We use KNN-based background subtraction to identify movement/disturbance in the video and use contours to get bounding-boxes for each item on the background.
-
We custom-made an Object Tracking algorithm to associate/group these bounding boxes into individual moving objects. We iterate over each frame and for every bounding-box found, we find the nearest, already-existing moving_object. If the found object and the box are both the nearest entities to each-other, they are associated/grouped into the same moving_object. If the box is left alone, it is created as a new moving_object. If a moving_object is left alone(for a few frames), we stop tracking that object for the rest of the video.
-
A few filters are applied on the moving objects to refine the output. These objects are now cropped out of the original video and overlaid on the background simultaneously (irrespective of their starting time in the original video) to give an appearance of all these events happening at the same time. Each event's time of occurrence is displayed on top of its moving object. The events are given a slight transparent appearance to deal with overlaps between multiple objects.
- Overlapping of multiple objects: One of the earliest problems we came across was that of multiple objects of different timelines overlapping with each other. We tried delaying the overlapping objects by some duration but that didn't cover all the cases. We ended up giving the moving objects an opacity of 0.5 so that overlapping objects didn't obstruct each other and since they're translucent, one was visible behind the other.
- Object tracking: Background subtraction only gives us boxes around people/things in each frame. Object tracking is the task of associating/grouping the boxes across multiple frames together. The object tracking solutions already implemented online were not in line with our purpose of usage, so we had to custom-implement an object tracking algorithm ourselves, the working of which is described above.
-
Download
$ git clone git@github.com:eonr/cctv-recap.git
-
The program is built and tested on
Python 3.7 -
Install dependencies
$ pip install -r requirements.txt
-
Run
$ python cctv_recap.py <path to video>
CCTV Recap is licensed under BSD-3-Clause.