Data Collection tools that use video: rules, laws, ethics...and data formats?

2 posts / 0 new
Last post
elizabethsall
Data Collection tools that use video: rules, laws, ethics...and data formats?

I have come across a few interesting use cases for computer vision in the past few days:

How many people are wearing masks vs not? [ in this particular case, when boarding a transit vehicle ]

I learned of a transit agency using video that was already on buses and running it through open source computer vision tools in house to tally the compliance rate.  I also learned that there were several packages that have been developed specifically for the purpose of "mask identification".

Ped/Bike counting from existing traffic video streams in real time

This open source package from LADOT (https://github.com/CityOfLosAngeles/automated-walk-bike-counterimplements algorithms described in this paper (https://www.semanticscholar.org/paper/An-End-to-End-Traffic-Vision-and-Counting-System-in-Wang-Owens/c1d98fca75c63fd5975fc2fcd3fe07ac02de4a5b?p2df) to process pedestrian and bike flows in real-time using GPU Tensor Flow processing. 

In addition to letting the community know about these interesting cases, I had the following questions:

1. Are there existing references summarizing "best practices" or laws related to processing, sharing, and storing video data? (particularly of people's faces)

As computer vision becomes increasingly accessible, and doable, these will be of increasing use.

2. I. know there were a lot of meetings and collaborations regarding pedestrian and bicycle data formats/standards for counts.  Are those projects still active or are there existing data specs for pedestrians and cyclists that have become a defacto standard?

3. Computer vision has (often rightfully) been criticized for putting product before ethics...but I'm particularly interested in how it could be used on the right side of this equation by potentially understanding racial and gender disparities in exposure rates.  Is anybody aware of any work that has (successfully and ethically) executed this?

And - there may be really obvious answers here...this is a new area for me.

Best,

Elizabeth Sall

qolilowoj

Hi Elizabeth,

Thanks for your interesting questions. As I have some experiences
working with computer vision experts, the responses below might be
useful.

1. In computer vision, it is not difficult to detect the face masks
using some deep learning frameworks, such as face detection, face
landmark detection. For example:
https://www.pyimagesearch.com/2020/05/04/covid-19-face-mask-detector-wit...

Processing, sharing and storing video data need a large amount of
memory and other computational resources. One commonly-used 30 fps
video feed generates 30 frames/images for 1 second. Normally, within a
GPU setting, it requires your AI algorithm to reach at least15 fps, in
order to meet real-time video processing requirements.

2. There are many pedestrian benchmark datasets available, but they
are mainly used in academic research communities for addressing
various challenges, e.g., for improving the performance and the
accuracy with human occlusions. Regarding data specs specifically for
pedestrians and cyclists, you can see the example below:
http://www.vision.caltech.edu/Image_Datasets/CaltechPedestrians/

3. How to integrate privacy-aware algorithms into computer vision
framework to protect privacy is still a new research area with very
limited applications in a city setting.

--
Xuesong (Simon) Zhou
School of Sustainable Engineering and the Built Environment
Ira A. Fulton Schools of Engineering
Arizona State University
Email: xzhou74@asu.edu

Xuesong

On Tue, Jun 23, 2020 at 12:59 PM elizabethsall wrote:
>
> I have come across a few interesting use cases for computer vision in the past few days:
>
> How many people are wearing masks vs not? [ in this particular case, when boarding a transit vehicle ]
>
> I learned of a transit agency using video that was already on buses and running it through open source computer vision tools in house to tally the compliance rate. I also learned that there were several packages that have been developed specifically for the purpose of "mask identification".
>
> Ped/Bike counting from existing traffic video streams in real time
>
> This open source package from LADOT (https://github.com/CityOfLosAngeles/automated-walk-bike-counter) implements algorithms described in this paper (https://www.semanticscholar.org/paper/An-End-to-End-Traffic-Vision-and-C...) to process pedestrian and bike flows in real-time using GPU Tensor Flow processing.
>
> In addition to letting the community know about these interesting cases, I had the following questions:
>
> 1. Are there existing references summarizing "best practices" or laws related to processing, sharing, and storing video data? (particularly of people's faces)
>
> As computer vision becomes increasingly accessible, and doable, these will be of increasing use.
>
> 2. I. know there were a lot of meetings and collaborations regarding pedestrian and bicycle data formats/standards for counts. Are those projects still active or are there existing data specs for pedestrians and cyclists that have become a defacto standard?
>
> 3. Computer vision has (often rightfully) been criticized for putting product before ethics...but I'm particularly interested in how it could be used on the right side of this equation by potentially understanding racial and gender disparities in exposure rates. Is anybody aware of any work that has (successfully and ethically) executed this?
>
> And - there may be really obvious answers here...this is a new area for me.
>
> Best,
>
> Elizabeth Sall
>
> --
> Full post: https://tmip.org/content/data-collection-tools-use-video-rules-laws-ethi...
> Manage my subscriptions: https://tmip.org/mailinglist
> Stop emails for this post: https://tmip.org/mailinglist/unsubscribe/13448