<?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 Mannat Dhingra on Medium]]></title>
        <description><![CDATA[Stories by Mannat Dhingra on Medium]]></description>
        <link>https://medium.com/@dhingra.mannat2007?source=rss-52c53c8b2b86------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*yJq9HJnmntu5oqiwo_NQDQ.jpeg</url>
            <title>Stories by Mannat Dhingra on Medium</title>
            <link>https://medium.com/@dhingra.mannat2007?source=rss-52c53c8b2b86------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Mon, 16 Nov 2026 22:49:52 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@dhingra.mannat2007/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[The Scarecrow Project: How a Small Computer Vision Challenge Became Something Much Bigger]]></title>
            <link>https://medium.com/@dhingra.mannat2007/the-scarecrow-project-how-a-small-computer-vision-challenge-became-something-much-bigger-c76f688d903c?source=rss-52c53c8b2b86------2</link>
            <guid isPermaLink="false">https://medium.com/p/c76f688d903c</guid>
            <category><![CDATA[retail-technology]]></category>
            <category><![CDATA[ai-for-business]]></category>
            <category><![CDATA[ambient-accountability]]></category>
            <category><![CDATA[yolov8]]></category>
            <category><![CDATA[computer-vision]]></category>
            <dc:creator><![CDATA[Mannat Dhingra]]></dc:creator>
            <pubDate>Sun, 19 Jul 2026 08:41:44 GMT</pubDate>
            <atom:updated>2026-07-19T08:41:44.061Z</atom:updated>
            <content:encoded><![CDATA[<p>The Scarecrow Project began with a simple challenge from a major retailer: could computer vision be used to decrease shrink by discourage theft before it happens?</p><p>The idea was not to identify someone as a thief or take action against them. It was simply to make people aware that they were being actively detected.</p><p>A camera at the entrance would detect a human, draw a box around them and display the processed live feed on a screen visible inside the store.</p><p>A genuine customer would probably continue normally. But someone entering with the intention of stealing might become more cautious after seeing themselves being detected.</p><p>That is where the scarecrow analogy came from.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*V-Pl0nbEsFkoxU_CsGgBxQ.png" /><figcaption>Scarecrow image</figcaption></figure><p>A scarecrow does not chase birds or physically stop them. Its visible presence creates uncertainty and discourages them from approaching the crop.</p><p>The Scarecrow Project applies a similar idea in a retail environment.</p><p>It does not accuse, confront or automatically judge anyone. It creates a visible deterrence effect by showing that the environment is actively monitored.</p><h3>Turning the Idea into a Complete Product Loop</h3><p>Building the human-detection model was only one part of the challenge.</p><p>For the solution to work in a real store, it needed to be easy to install and configure without changing the code for every location.</p><h3>Connecting to a Camera</h3><p>The first step was connecting the application to a camera.</p><p>I built a camera-scanning feature that searches the local Wi-Fi network and lists the available camera feeds. The user can then select and configure the camera required for the store.</p><blockquote><strong><em>Under the hood:</em></strong><em> IP cameras on a store network speak RTSP — a streaming protocol that lets the app pull a live video feed directly over Wi-Fi. One practical problem: most cameras get a different IP address every time they reconnect to the network (DHCP). To handle this reliably, the scanner looks up each camera by its fixed hardware MAC address using an ARP lookup, so the right camera is always found regardless of what IP it was assigned that day. Once discovered, the user enters the RTSP credentials once and the app remembers the connection.</em></blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/900/1*wAptOlcgb6Sdh35tJwzTaw.png" /></figure><h3>The Detection Engine</h3><p>Once the feed starts, the computer vision engine processes the video in real time, detects a human and draws bounding boxes around them.</p><blockquote><strong><em>Under the hood:</em></strong><em> Detection runs on </em><strong><em>YOLOv8</em></strong><em>, a state-of-the-art real-time object detection model. But detection alone only tells you </em>where<em> a person is in a single frame — not whether it is the same person from the previous frame. For that, the engine uses </em><strong><em>ByteTrack</em></strong><em>, a multi-object tracking algorithm that assigns each detected person a stable ID across frames. This is what allows the system to count individuals and keep a consistent bounding box on someone as they move through the frame. To keep the feed smooth and non-lagging, a dedicated background thread continuously grabs the newest camera frame — the detection loop always works on fresh footage, never a stale queue.</em></blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/900/1*zgq1xow1MI2cAoksYe0B-g.png" /></figure><h3>Casting to the Store Screen</h3><p>The next challenge was displaying the processed feed on a television or digital screen inside the store.</p><p>To solve this, I built another scanning feature that discovers cast-enabled devices available on the same Wi-Fi network. The user can select the required display, and the processed feed can then be cast directly from the application.</p><blockquote><strong><em>Under the hood:</em></strong><em> The app uses </em><strong><em>pychromecast</em></strong><em> — a Python library that speaks the Google Cast protocol — to discover Chromecast devices and Google TVs on the local network automatically. The processed live feed (complete with detection boxes) is served as an MJPEG stream by a built-in </em><strong><em>Flask</em></strong><em> web server and pushed to the selected screen. No HDMI cables, no separate casting device, no manual IP entry — the app handles the full handshake. Other protocols related to Apple/Samsung/Other digital devices are work in progress.</em></blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/900/1*pziOoWyxMBAtsxoj0zOoxg.png" /></figure><p>This creates the complete deterrence loop:</p><p><strong>Discover the camera → Configure it → Start the feed → Detect people → Discover the screen → Cast the processed feed → Create visible deterrence</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*xDnmd9LTTLrZr4YCYkA1CA.png" /></figure><h3>Why the Complete Loop Matters</h3><p>A working computer vision model is not automatically a usable product.</p><p>A store should not need a developer every time it wants to connect a new camera or change the display screen. The application must discover devices, allow configuration and manage the complete workflow through the product itself.</p><blockquote><strong><em>Architecturally</em></strong><em>, this meant building a Flask-based web application where every setting — camera credentials, detection thresholds, cast targets — is configurable through a browser UI and stored in a config file. No Python file needs to be edited. The same codebase deploys to a different store by opening the browser, not by touching the code.</em></blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/900/1*VrPty2AvMVaIdF5glqj9lQ.png" /></figure><p>This project helped me understand that building a product is about more than writing the detection code.</p><p>It is also about solving everything around the model:</p><ul><li>How the camera is discovered</li><li>How the user selects the right feed</li><li>How the detection engine is activated</li><li>How the processed output reaches a screen</li><li>How the entire setup can be repeated in another location</li></ul><p>The technology becomes useful only when the full experience works from beginning to end.</p><h3>Beyond the Store — The Productivity Case</h3><p>The deterrence principle at the heart of this project does not begin and end at a retail entrance.</p><p>In any environment where consistent human performance matters, the most persistent challenge is rarely capability. It is consistency — the gap between what people are capable of and what they actually deliver when accountability feels invisible.</p><p>There is a well-documented phenomenon called the <strong>Hawthorne effect </strong>— the observation that people tend to perform closer to their best simply when they are aware they are being observed. A visible monitoring system does not need to record, judge, or intervene in real time to produce this effect. The screen showing active detection is sufficient. The knowledge that the environment is paying attention tends to raise standards on its own — not out of fear, but because accountability becomes part of the atmosphere rather than something imposed through periodic reviews.</p><blockquote>The scarecrow does not need to move. It only needs to be seen.</blockquote><p>This applies as naturally to a warehouse loading bay as it does to a store entrance. A processing unit where staff can see that activity is being tracked tends to operate with more consistent focus. A customer service floor where engagement is visibly monitored tends to self-regulate more effectively. A back-office team working in an environment with active monitoring in place tends to apply more discipline to how time is used.</p><p>The intent is not surveillance. It is <strong>ambient accountability</strong> — the kind that is built into the space rather than enforced through interventions after the fact.</p><h3>What Comes Next</h3><p>The current Scarecrow use case focuses on visible deterrence through real-time human detection and casting.</p><p>While building it, I started exploring several other possibilities, including face recognition, zone density, PPE detection, multi-camera operations, loitering detection, event logging and rule-based alerts.</p><p>I will share those features separately in future posts.</p><p>For now, the most important learning from the Scarecrow Project is simple:</p><blockquote>Sometimes the most effective technology does not need to take action. Its visible presence may be enough to influence behaviour.</blockquote><p>What started as a box around a person became an exercise in computer vision, product thinking and building a complete end-to-end experience.</p><h3>Pilot Program — Try It Before It Goes Mainstream</h3><p>The product is working. The architecture is stable. What it needs now is real-world context across different industries and environments.</p><p>If you are in retail, warehousing, back-office operations, customer service, or any setting where consistent human performance is both important and difficult to sustain, I would like to offer you a free pilot.</p><p>There is no cost and no obligation. The goal is to deploy the system in a real environment, understand what works, gather honest feedback, and use that to build something genuinely useful before it scales.</p><p>In return, you get a working computer vision monitoring system — set up, configured, and supported for the duration of the pilot — with full transparency into what it is doing and why.</p><p>Fill in the short form below to express your interest, and I will follow up personally. Or reach out directly via LinkedIn if you would prefer to start a conversation first.</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fdocs.google.com%2Fforms%2Fd%2Fe%2F1FAIpQLSc7rZaV8UY20Wjbk9Ltjdxwm4WU_uoZbp0NmsV0q9iMgPWHCQ%2Fviewform%3Fembedded%3Dtrue&amp;display_name=Google+Docs&amp;url=https%3A%2F%2Fdocs.google.com%2Fforms%2Fd%2Fe%2F1FAIpQLSc7rZaV8UY20Wjbk9Ltjdxwm4WU_uoZbp0NmsV0q9iMgPWHCQ%2Fviewform%3Fusp%3Dsend_form&amp;image=https%3A%2F%2Flh3.googleusercontent.com%2F38p8O4faSxQdilFjpvxDumoaOqe-E745R3gSN_c3ovbQ9IS6-MTavfsBK75w_NhN2LzbYyAro8SfOTE%3Dw1200-h630-p&amp;type=text%2Fhtml&amp;scroll=auto&amp;schema=google" width="760" height="675" frameborder="0" scrolling="no"><a href="https://medium.com/media/a45859c5b3d0e44189135e1bc2f5f3c9/href">https://medium.com/media/a45859c5b3d0e44189135e1bc2f5f3c9/href</a></iframe><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c76f688d903c" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>