[Academics] VISTA: A System of Filtering Spatial Content of Videos for Traffic Surveillance(10/25)
- College of Software and Engineering
- Hit710
- 2022-10-24
- Title: VISTA: A System of Filtering Spatial Content of Videos for Traffic Surveillance
- Speaker: Arani Bhattacharya
Indraprastha Institute of Information Technology Delhi, India
- room : Engineering 2 building 26106 17:00 ~ 18:30
Traffic surveillance via deployed cameras on roads is
increasingly used to prevent accidents and improve traffic
security. Such traffic cameras send video feeds to a cloud
or edge center to run computer vision algorithms, which is
bandwidth-intensive. Current state-of-the-art run DNNs
extract a limited number of frames/pixels/regions or rely
on re-encoding the important parts of the video. This
imposes the overhead of re-encoding on the camera side,
which can be significant. In this work, we propose
Tile-Filter, a system that utilizes tile sampling, where
a limited number of rectangular areas within the frames,
known as tiles, are sent to the server. We then propose
an adaptive tile filtering algorithm that estimates the
presence of moving objects by comparing the statistics of
the tiles' bitrate (in kbps) to retain only the necessary
tiles. We evaluate VISTA on a dataset of 56 videos to show
that on average our technique reduces 17-40\% of the total
amount of data sent to the cloud while providing a detection
accuracy of over 0.85. Furthermore, VISTA also runs in
real-time even on cheap edge devices like Raspberry Pi and
nVidia Jetson Nano. Further, it requires minimal calibration
compared to prior works.