<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Niladri Kumar Sahoo on Medium]]></title>
        <description><![CDATA[Stories by Niladri Kumar Sahoo on Medium]]></description>
        <link>https://medium.com/@niladrikumarsahoo.26?source=rss-8e81d19fc6fd------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*UAYsMv56oLhai_-v7i6zpw.png</url>
            <title>Stories by Niladri Kumar Sahoo on Medium</title>
            <link>https://medium.com/@niladrikumarsahoo.26?source=rss-8e81d19fc6fd------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sat, 21 Nov 2026 10:06:01 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@niladrikumarsahoo.26/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Can AI See a Stampede Coming? Building a Crowd-Monitoring Prototype.]]></title>
            <link>https://medium.com/@niladrikumarsahoo.26/can-ai-see-a-stampede-coming-building-a-crowd-monitoring-prototype-e06bf4e0e05e?source=rss-8e81d19fc6fd------2</link>
            <guid isPermaLink="false">https://medium.com/p/e06bf4e0e05e</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[computer-vision]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[python]]></category>
            <category><![CDATA[research]]></category>
            <dc:creator><![CDATA[Niladri Kumar Sahoo]]></dc:creator>
            <pubDate>Mon, 26 Oct 2026 13:27:02 GMT</pubDate>
            <atom:updated>2026-10-26T13:27:02.675Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*8KBweWjBlB7RV_Ew9u2B8Q.jpeg" /><figcaption>Non-calibrated crowd stampede warning system taken from prototype. (Fig. 1)</figcaption></figure><p>Residing in the cultural capital of India, <strong>Kolkata</strong>, I know how challenging it is to get in and out of large crowds, temporary pandals, traffic and blockades, temporary large open-sided pavilions, during the festival or an event. Around<strong> </strong><a href="https://dataful.in/collections/796/"><strong>145 to 250+</strong> <strong>people are a part of stampedes</strong></a> at a specific site or event, often exceeding local venue or official capacity limits every year.</p><h4>The idea behind Crowd-Pilot AI</h4><p>The basic concept is quite straightforward.</p><p>A camera captures a crowd. A computer vision model estimates the number and distribution of people. The system divides the monitored area into zones and uses the estimates to identify places that might need attention.</p><p>Eventually, I want Crowd-Pilot AI to help festival organizers and authorities understand where crowds are gathering and where additional monitoring or intervention might be needed.</p><p>But before thinking about dashboards, alerts, or routing, I needed to solve a more fundamental problem.</p><p><strong>Can a model count the people who are actually inside the area we care about?</strong></p><h3>My first working pipeline</h3><p>I built an initial pipeline using Python, OpenCV, and CSRNet, a pretrained deep-learning model for crowd counting.</p><p>The pipeline takes a video frame, estimates crowd density, and produces a visual representation of where people are concentrated. I also experimented with dividing the monitored floor into a grid to estimate the number of people in individual zones.</p><p>The goal was to move beyond a single crowd count.</p><p>Knowing that a scene contains hundreds of people is useful, but it doesn’t tell an organizer whether one particular entrance, corridor, or section of a venue is becoming overcrowded.</p><p>A zone-based approach could eventually provide more actionable information.</p><p>To test the pipeline, I used a crowd video available on the web. The footage was not from a Kolkata festival, and I have not validated the model’s accuracy on Indian festival crowds.</p><p><strong>Still, the experiment exposed a problem worth investigating.</strong></p><h3>The 361-versus-211 problem (Fig. 1)</h3><p>In one frame, the model’s estimated count was approximately 361 people when considering the entire visible scene.</p><p>But the area I actually wanted to monitor was the hall floor outlined in my visualization.</p><p>When I restricted the analysis to that region, the estimate was approximately 211 people.</p><p>The difference was significant.</p><p><strong>361 people estimated across the scene. Approximately 211 estimated within the monitored floor region.</strong></p><p>The scene included a crowd outside the glass doors. Those people were visible to the model, but they weren’t part of the floor area I was trying to monitor. This distinction matters because a crowd-monitoring system shouldn’t treat every visible person as part of the same monitored space. A person standing outside an entrance, someone inside a corridor, and someone in a crowded section of a hall may require entirely different interpretations depending on the area being monitored.</p><p>The model was producing a count, but the application needed something more specific: <strong>a count associated with a meaningful physical region.</strong></p><p><strong>That became the most useful lesson from this experiment.</strong></p><h3>Why a headcount isn’t the same as crowd density.</h3><p>Another issue became clear as I worked on the prototype.</p><p>Counting people is <strong>not the same</strong> as measuring crowd density.</p><p>A count of 50 people could represent a relatively spread-out group covering a large area or a tightly packed group in a much smaller space.</p><p>For crowd safety, the number of people per unit area is often more informative than the total count alone.</p><p>But converting an image-based estimate into people per square meter is not as simple as dividing the count by the number of pixels inside a boundary.</p><p>Perspective changes how physical distances appear in an image. A section of floor near a camera can occupy far more pixels than a physically similar section farther away. Camera angle, lens characteristics, and the geometry of the monitored space all affect the relationship between image coordinates and real-world distances.</p><p>A useful system therefore needs camera-specific calibration and a way to map image regions to physical areas.</p><p>I haven’t implemented or validated that calibration yet.</p><p>Until then, the prototype’s outputs should be treated as experimental crowd-count estimates, not reliable measurements of physical density or indicators of immediate danger.</p><h3>What I want to work on next</h3><p>There are three main areas I want to explore.</p><p><strong>1. Calibrating the monitored floor</strong></p><p>My first priority is to establish a more reliable relationship between image regions and physical space. This includes investigating <strong>camera geometry, perspective, and how to convert estimates into meaningful density measurements.</strong></p><p><strong>2. Testing on more relevant footage</strong></p><p>The pipeline needs testing across different scenes, camera angles, lighting conditions, crowd sizes, and movement patterns. Footage of Indian festival crowds would be particularly relevant to the original motivation for this project.</p><p><strong>3. Building a dashboard for organizers</strong></p><p>Once the underlying estimates are more reliable, I want to explore a dashboard that displays crowd estimates by zone and helps organizers identify areas that deserve closer attention.</p><p>Any alert logic would need careful validation, and the system should <strong>support human decision-making</strong> rather than claim to determine safety on its own.</p><p>Access to suitable footage is another challenge. The current prototype uses publicly available web footage. I don’t have access to live police camera feeds. Working with real operational footage would require appropriate permission and privacy safeguards.</p><h3>Why I’m Building This</h3><p>This project started from one motivation, to do something that can help those around me, with my project, Crowd-Pilot AI.</p><p>Sometimes I wasn’t sure if this project would be of any use. But now I feel more inclined to finish it.</p><p>Durga Puja in Kolkata is more than a festival. Artistic Pandals/Pavilions are built all over West Bengal for Durga goddess’s worship. People travel through the crowded street, narrow lane, and temporary footpath to view these pandals that draw huge crowds.</p><p>However, handling the crowds at this scale presents its own set of <strong>problems: roads are clogged, routes no longer look like the ones you recognize, and visitors arrive only to find that the road they want to take just opened, but it’s now closed. </strong>In tightly packed places, access can be problematic without information.</p><p>Bad crowd control can have deadly results. Recent reports of tragedies at festivals serve as a reminder that crowd safety is not about the number of people in the crowd, it’s about lighting, access paths, means of communication, plans in place for emergencies and intervening swiftly.</p><p>That gave me some motivation for thought about what I could bring to the table as a human learning computer vision and honing AI projects while I build this project.</p><p><strong>What if a system could: Identify potentially dangerous concentrations of crowds early, alert people to changing routes, and give organizers better information about where they might want to focus their efforts?</strong></p><p>So that’s what prompted me to start up Crowd-Pilot AI.</p><p>My aim is to investigate two related functionalities: the ability to offer assistance to visitors to make more informed decisions about navigating streets and avoiding crowded zones, and the ability to inform organisers and authorities at the zone level that crowd situations might become worse.</p><p>Both features are not addressed by my current prototype. Accurate density estimation, full-fledged field tests, data, and being able to work with the crowd managers would be necessary before such a system could be deployed in an operational setting.</p><p><strong>I am also aware that AI cannot substitute trained staff, emergency planning, physical facilities or human judgement.</strong></p><p><strong>But I think computer vision will genuinely be worth exploring as an effort towards crowd safety.</strong></p><p>Crowd-Pilot AI is so new that I don’t know how far I can go with it, nor what the technical solutions are that are ready for the real world. What I do know is the reason I began.</p><p>I would love to see if anything I create has the potential to assist and/or guide crowds through large public gatherings more safely.</p><p>That’s what makes this project worth continuing for me.</p><p><strong>Link to the project: </strong><a href="https://github.com/NiladriKrSahoo-dev/Crowd-pilot-AI">https://github.com/NiladriKrSahoo-dev/Crowd-pilot-AI</a></p><p><strong><em>Crowd Pilot AI is an experimental prototype under development. It has not been validated for operational crowd-safety use.</em></strong></p><p><strong><em>I’m particularly interested in feedback from people working in computer vision, crowd dynamics, public safety, or video analytics. I’d like to learn more about spatial calibration, ground-truth evaluation, and the challenges of applying crowd-counting models to real festival environments.</em></strong></p><p><strong><em>The footage is from the web and isn’t mine. Happy to credit the owner or remove it on request.</em></strong></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e06bf4e0e05e" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>