<?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 Mehdi Sqalli on Medium]]></title>
        <description><![CDATA[Stories by Mehdi Sqalli on Medium]]></description>
        <link>https://medium.com/@msqalli?source=rss-151f42ef18b8------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*xCWgdbEzHEra7BvzHL78RQ.png</url>
            <title>Stories by Mehdi Sqalli on Medium</title>
            <link>https://medium.com/@msqalli?source=rss-151f42ef18b8------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Fri, 20 Nov 2026 22:28:24 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@msqalli/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[Small, Honest Software in the Age of AI]]></title>
            <link>https://msqalli.medium.com/small-honest-software-in-the-age-of-ai-cc00968a3faa?source=rss-151f42ef18b8------2</link>
            <guid isPermaLink="false">https://medium.com/p/cc00968a3faa</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[software-engineering]]></category>
            <category><![CDATA[ios-development]]></category>
            <category><![CDATA[startup]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Mehdi Sqalli]]></dc:creator>
            <pubDate>Sat, 28 Feb 2026 00:58:16 GMT</pubDate>
            <atom:updated>2026-02-28T00:58:16.526Z</atom:updated>
            <content:encoded><![CDATA[<p>I’ve spent years building complex systems.</p><p>Startups. Distributed systems. Marketplaces. Maps. AI-powered products. Software that requires, roadmaps, trade-offs, meetings, and constant iteration.</p><p>I enjoy building ambitious things.</p><p>But lately, in my spare time, I’ve been building something very different.</p><p>Small, honest software.</p><p>And I would not have been able to do it without AI.</p><h3>I’m Tired of “Simple” Apps That Aren’t Simple</h3><p>Over time, something started to bother me.</p><p>Very simple apps asking for subscriptions.<br> Tiny utilities collecting data they don’t need.<br> Free apps filled with ads and tracking SDKs.</p><p>A habit tracker that requires an account / subscription.<br>A tiny puzzle game filled with ads.<br>A white noise app collecting analytics / needing a subscription.</p><p>Software that should be small and respectful has become complicated and extractive by default.</p><p>So I started asking myself what it would look like to build extremely simple apps that:</p><ul><li>work fully offline</li><li>collect zero data by design</li><li>have no ads</li><li>have no subscriptions</li><li>don’t require accounts</li><li>don’t track anything</li></ul><p>Just small tools. You download them. They work. They are yours. That’s it.</p><h3>AI Changed the Equation</h3><p>Two years ago, this would have been unrealistic for me.</p><p>Even with experience, building and polishing a small iOS app properly takes time. More than people think.</p><p>A crossword app? That’s weeks/months.<br>A simple sounds app? Also weeks.</p><p>Not because the ideas are complicated, but because implementation always is. Edge cases, UI polish, state management, bugs, App Store submission, metadata, refactoring, QA etc etc.</p><p>And to be clear, I understand why many apps rely on subscriptions, ads, analytics, and accounts. Servers cost money. Support costs time. Teams need to get paid. For many products, all those mechanisms obviously make sense.</p><p>This isn’t a moral judgment.</p><p>It’s just that for very small, self-contained tools, I wanted to see if there was another way.</p><p>And Now, with AI, the equation is different.</p><p>I mostly architect and guide. I describe what I want. I review. I refactor. I adjust. Less and less to be honest :)</p><p>The crossword app that would have taken me weeks took me a weekend.</p><p>The white noise app that would have taken a few weeks took me a few hours.</p><p>Over time, I ended up shipping more than 10 small iOS apps.</p><p>That simply would not have happened before.</p><h3>Why They Cost $1</h3><p>All of these apps cost $1.</p><p>Not because I’m trying to maximize revenue. And not because I expect thousands of reviews or scale.</p><p>It’s more about the contract.</p><p>When someone pays even a dollar, something changes. There’s intent. There’s a small but real commitment.</p><p>If a user takes the time to leave feedback, I know they cared enough to pay for it. I take that seriously. I trust that signal more.</p><p>It also keeps things simple. No ads. No upsells. No subscription logic.</p><p>You pay once. You get the app. It works offline. It’s YOUR app. That’s the entire deal.</p><h3>Offline by Design</h3><p>All the apps work fully offline.</p><p>There are no analytics SDKs.<br> No ad frameworks.<br> No accounts.<br> No backend.</p><p>I’m not trying to make a statement. These apps simply don’t need any of that.</p><p>By removing the backend entirely, I remove the temptation to collect anything. There is nothing to optimize, nothing to track, nothing to monetize beyond the initial dollar.</p><p>It forces simplicity.</p><p>And I find that simplicity refreshing.</p><h3>This Is Not a Startup</h3><p>I’m not trying to build the next big thing.</p><p>I’m not raising money.<br>I’m not optimizing funnels.<br>I’m not trying to turn this into a company.</p><p>So far, these small apps have reached 1300+ downloads.</p><p>That’s not massive. Many people are building far more impressive things.</p><p>But it’s real.</p><p>Real people downloaded something small that does one thing and respects them.</p><p>Most of these apps are built in the margins. An hour here. A weekend there.</p><p>Without AI, they would have remained ideas.</p><p>I’ve been publishing these small experiments under a simple umbrella called <a href="https://localonelabs.com/">LocalOneLabs</a>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*OQUiJCMvV7cKMtDrZrbnpg.png" /></figure><h3>AI Is Already Changing How We Build</h3><p>There are a lot of loud takes about AI right now.</p><p>Some people expect immediate revolutions. Others focus on the risks. Like any technology, it will bring both.</p><p>What feels undeniable to me is this: the barrier to building has shifted.</p><p>You can already see it in small ways everywhere.</p><p>For me, it meant I could finally act on ideas that were too small to justify weeks of work.</p><p>AI doesn’t just make startups faster.</p><p>It allows individuals to build things that would previously have required much more time or a team.</p><p>And that opens up infinite possibilities.</p><h3>A Small Way of Giving Back</h3><p>I’ve been lucky in my career. I’ve had the chance to work on complex systems and ambitious products.</p><p>This feels different.</p><p>It’s not about scale. It’s not about growth.</p><p>It’s about building small tools that respect the people who use them.</p><p>Software can be straightforward.<br>It can run entirely on your device.<br>It can cost a dollar and nothing more.</p><p>And it can still be worth building.</p><p>If you’re an engineer who cares about craft, simplicity, and building things that treat users fairly, this feels like a fascinating time to experiment.</p><p>AI makes it possible to build faster.</p><p>But more importantly, it gives us the option to build differently.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=cc00968a3faa" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Vehicle detection]]></title>
            <link>https://msqalli.medium.com/vehicle-detection-497a779fbe21?source=rss-151f42ef18b8------2</link>
            <guid isPermaLink="false">https://medium.com/p/497a779fbe21</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[self-driving-cars]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[udacity]]></category>
            <category><![CDATA[computer-vision]]></category>
            <dc:creator><![CDATA[Mehdi Sqalli]]></dc:creator>
            <pubDate>Mon, 27 Mar 2017 03:51:38 GMT</pubDate>
            <atom:updated>2017-03-27T03:51:38.692Z</atom:updated>
            <content:encoded><![CDATA[<p><em>You can find my code </em><a href="https://github.com/MehdiSv/VehicleDetection"><em>here</em></a><em>. It might be more up to date than the article ;).</em></p><h3>Intro</h3><p>This is the fifth article in the self driving cars series. If you want to know why I’m sharing this and more about my journey, please read <a href="https://medium.com/@MSqalli/self-driving-car-journey-to-build-one-5413ad0e4115">this</a>.</p><h3>Goal</h3><p>For each image of the given video, we are going to identify cars and draw a bounding box around them.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/640/1*gM1OA3x_FGq_UtIJN_yjNQ.png" /><figcaption>Vehicle detection final result</figcaption></figure><h3>Algorithm: Step by step</h3><p>This algorithm is executed step by step on each frame:</p><p>1: HOG features extraction for learning</p><p>I first computed the HOG of many car and non car images.</p><p>HOG, or Histogram of Oriented Gradients is a technique that maps all the location of the gradients (change in color or luminosity) in a picture, as well as their orientation.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/64/1*tZpSjE7lmuLEUxpIZHTD2w.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/64/1*MBhyCV3YM3MEz_C-ZuDF5g.png" /><figcaption>car and non car image</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/640/1*QkF8q0-R8eJaLWPpQdoXBA.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/640/1*APPZ_iKyNk9u-aKT0-GwBA.png" /><figcaption>car and non car HOG</figcaption></figure><p>2: Spatial binning</p><p>We can perform spatial binning on an image and still retain enough information to help in finding vehicles.</p><p>This allows us to filter out some of the noise and only retain the most important features of the car, while making the algorithm run faster as the resulting images are smaller.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/839/1*5lzYsorWLinmPmC11AIYjw.jpeg" /><figcaption>Spatial binning example</figcaption></figure><p>3: Color channel histograms</p><p>This will give us pixel intensity for each image, allowing us to use the common car colors as a feature for recognition.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ECW9QPVl-oaSI_6cGjkkYA.jpeg" /><figcaption>color histogram example</figcaption></figure><p>4: Training a classifier</p><p>With these 3 features (HOG, spatial bins, color histograms) we train a classifier to separate cars from the rest.</p><p>I did the training using a pretty standard method:</p><ul><li>I first standardised the computed features using a standard scaler</li><li>Then I split my data into training and test sets (80/20), randomising them in the process.</li><li>I finally trained a standard SVC with default settings.</li></ul><p>I also stored the results of the computed features, scaler and SVC to pickle files to avoid having to recompute them every time.</p><p>5: Final choice of parameters</p><p>Using the accuracy of the classifier, I tweaked the parameters to obtain a better result.</p><p>For the HOG, changing the number of orientations, of pixels per cell and cells per block never gave significantly better results. So I mainly tweaked the color space and channel used, in particular RGB (red or all channels), HSV (every combination) and YCrCb (all channels).</p><p><strong>YCrCb on all channels gave the best results, with the original HOG settings: 8 orientations, 8*8 pixels per cell and 2*2 cells per block.</strong></p><p>6: Sliding window</p><p>I then used a sliding window approach to search for cars in the image.</p><p>I restricted the windows to the lower half of the image. I also added different window sizes, ranging from 64*64 to 128*128.</p><p>Based on the size of the window, I also further restricted their location. For example, 64*64 windows are only computed on the top half of the lower half of the image as cars located on the lowest part of the image are supposed to be up close and thus appear bigger.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*8YBtMtyaE3zvMmApQ0ulXA.png" /><figcaption>sliding window</figcaption></figure><p>Finally, I made my window overlap a lot. We’ll see why in the next step.</p><p>7: Heatmap</p><p>The classifier not being 100% correct, we need to handle false positives.</p><p>I used a simple heatmap to do that. The overlapping windows allowed me to set a higher threshold on my heatmap which gave me significantly better results, detecting more true positives and less false positives.</p><p>By saving the detection results over the last five frames using a circular buffer and applying a threshold on the resulting heatmap, I get a result like this:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/640/1*YDAijBF8NAqQR5FtF8KSHA.png" /><figcaption>Final heatmap</figcaption></figure><p>Here is the full video:</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FbU-8mx7PV8E%3Ffeature%3Doembed&amp;url=http%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DbU-8mx7PV8E&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FbU-8mx7PV8E%2Fhqdefault.jpg&amp;key=d04bfffea46d4aeda930ec88cc64b87c&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/e0c2e417cacf27f48f3e0085ed8ca8fd/href">https://medium.com/media/e0c2e417cacf27f48f3e0085ed8ca8fd/href</a></iframe><h3>Thoughts</h3><p>Obviously what we did here is naive implementation.</p><p>The pipeline is likely to fail if :</p><ul><li>Worse weather/lighting conditions.</li><li>Less contrast between the cars and the road in general.</li><li>Going up/down hills.</li></ul><p>We used classical computer vision techniques and added a SVC classifier in this project. It’s really interesting to understand and get familiar with these concepts.</p><p>But again, I’d be very interested to see how deep learning would tackle this problem.</p><p>Next time, we’ll use Kalman filters to track the car in motion thanks to Radar and Lidar data! and all coded in C++ for real time execution!</p><p><strong>Stay tuned.</strong></p><p><em>Thanks to </em><a href="https://medium.com/u/d8603e818394"><em>Eric Sauvegrain</em></a> <em>for his feedback on this post.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=497a779fbe21" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Advanced Lane detection]]></title>
            <link>https://msqalli.medium.com/advanced-lane-detection-6a769de0d581?source=rss-151f42ef18b8------2</link>
            <guid isPermaLink="false">https://medium.com/p/6a769de0d581</guid>
            <category><![CDATA[computer-vision]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[self-driving-cars]]></category>
            <category><![CDATA[udacity]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Mehdi Sqalli]]></dc:creator>
            <pubDate>Mon, 13 Feb 2017 17:52:32 GMT</pubDate>
            <atom:updated>2017-02-13T17:52:32.833Z</atom:updated>
            <content:encoded><![CDATA[<p><em>You can find my code </em><a href="https://github.com/MehdiSv/AdvancedLineDetection"><em>here</em></a><em>. It might be more up to date than the article ;).</em></p><h3>Intro</h3><p>This is the fourth article in the self driving cars series. If you want to know why I’m sharing this and more about my journey, please read <a href="https://medium.com/@MSqalli/self-driving-car-journey-to-build-one-5413ad0e4115">this</a>.</p><h3>Goal</h3><p>For each image of the given video, we are going to detect the driving lane lines on each side of the car and visually highlight the area between them.</p><p><em>Please note: In a previous article I already described a </em><a href="https://medium.com/@MSqalli/lane-detection-446986c44021"><em>lane detection algorithm</em></a><em>. It did not work well on curved roads. I will handle that use case here.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ZRmqryQGeWSpmn37W6ToeQ.jpeg" /></figure><h3>Algorithm: Step by step</h3><p>This algorithm is executed step by step on each frame:</p><p>1: Camera Calibration and Image Undistortion.</p><p>When a picture is taken, 3D objects in the real world are transformed into a 2D image. This causes distortion.</p><p>The first step in analyzing camera images, is to undo this distortion.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ID3VkXrY4lRwrPZbVCz5jA.png" /></figure><p>Thanks to at least 20 chessboard images as shown above, we can correctly calibrate our camera and correctly generate our undistorted images.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*KgWQIyomkIlejrC5syukPw.jpeg" /></figure><p>2: Image filtering.</p><p>We use three different thresholds to filter the image.</p><ul><li>Color thresholding</li></ul><p>We convert our image to the HSV color space. It allows us to filter much more easily by focusing on the hue and saturation of the pixel (which color it is and how intense the color is), and not so much on its value (how dark it is), thus handling shadows and overall worse lighting conditions much more easily.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*BGO2zIgcN7PMv4yBiuzLpQ.jpeg" /></figure><ul><li>Direction threshold</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*15gl_F_4wmMzSHoXo-NiFQ.jpeg" /></figure><ul><li>Magnitude threshold</li></ul><p>In my <a href="https://medium.com/@MSqalli/lane-detection-446986c44021">previous lane detection algorithm</a>, I used the Canny Edge detection to detect edges. Here, I’ll use the Sobel operator in a slightly different manner.</p><p>The Sobel operator is at the heart of the Canny edge detection algorithm.</p><p>This will give us the ability to use a different threshold for the magnitude and the direction of the gradient.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*idKkWw7a-Hcx3M3ao-mNVw.png" /></figure><p>Once we combine the different thresholds previously computed we get clean lanes:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*zV2IE6rVg6S75g1mt6In1g.jpeg" /></figure><p>3: Perspective transform</p><p>A perspective transform maps the pixels in a given image to different ones with a new perspective. The perspective transform we’ll be using is a bird’s-eye view transform that let’s us view a lane from above.</p><p>This will help us calculate the lane curvature.</p><p>These are the points I used in my transformation:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/452/1*_NYw5hnlzC8-AJ0PhGvjZQ.png" /></figure><p>Which leads to this result:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*zFDr1dClHbQJYvKAL6zlcA.jpeg" /></figure><p>4: Lane detection</p><p>To detect the lanes I computed the histogram of the picture on its lower half to find the rough position of each lane line.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/800/1*ZsKR-EBCxQCzuHsXdWSNNw.png" /></figure><p>I then ran a sliding window vertically to detect the position of the center of each lane line on each part of the image.</p><p>After that I used these positions to compute polylines describing each lane line using a polyfit method.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ayVES5vkyOmuVLnFcCFGYg.jpeg" /></figure><p>5: Displaying the detected lane.</p><p>I then created a polygon using the curves of each computed polyline and warped back the result.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*-57BrIKJfj3m9lbT-2jZ8Q.jpeg" /></figure><p>Here is the full video:</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2F-Cb7HKFeJ5E%3Ffeature%3Doembed&amp;url=http%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3D-Cb7HKFeJ5E&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2F-Cb7HKFeJ5E%2Fhqdefault.jpg&amp;key=d04bfffea46d4aeda930ec88cc64b87c&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/c9337992cb9fb131f77e5631887e199e/href">https://medium.com/media/c9337992cb9fb131f77e5631887e199e/href</a></iframe><h3>Thoughts</h3><p>This project came after 2 deep learning projects. It’s really interesting to understand and get familiar with computer vision techniques.</p><p>Ultimately, this only made me appreciate deep learning even more. Deep Learning needs a lot more data, but is so much more robust. Once you accumulate enough data and fine tune well the hyper parameters, the network will be less impacted by lightning conditions and will scale better to adapt to new unseen environments.</p><p>Using computer vision techniques you ultimately fine tune your parameters to make it work in the environment you have seen.</p><p>Next time we’ll keep using computer vision techniques to detect vehicles! <strong>Stay tuned.</strong></p><p><em>Thanks to </em><a href="https://medium.com/u/d8603e818394"><em>Eric Sauvegrain</em></a><em>, </em><a href="https://medium.com/u/e027d96f8057">Aboud Jardaneh</a><em>, and </em><a href="https://medium.com/u/2087f6dc46">Salwa</a> <em>for their feedback on this post.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=6a769de0d581" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Teach a car how to drive with Deep Learning.]]></title>
            <link>https://msqalli.medium.com/teach-a-car-how-to-drive-b54e921f64d2?source=rss-151f42ef18b8------2</link>
            <guid isPermaLink="false">https://medium.com/p/b54e921f64d2</guid>
            <category><![CDATA[deep-learning]]></category>
            <category><![CDATA[self-driving-cars]]></category>
            <category><![CDATA[udacity]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Mehdi Sqalli]]></dc:creator>
            <pubDate>Mon, 30 Jan 2017 04:26:21 GMT</pubDate>
            <atom:updated>2017-02-04T17:53:41.570Z</atom:updated>
            <content:encoded><![CDATA[<p><em>You can find my code </em><a href="https://github.com/MehdiSv/BehavioralCloning"><em>here</em></a><em>. It might be more up to date than the article ;).</em></p><h3>Intro</h3><p>This is the third article in the self driving cars series. If you want to know why I’m sharing this and more about my journey, please read <a href="https://medium.com/@MSqalli/self-driving-car-journey-to-build-one-5413ad0e4115">this</a>.</p><h3>Goal</h3><p>Teach a virtual car how to drive on a closed circuit using <a href="https://devblogs.nvidia.com/parallelforall/deep-learning-self-driving-cars/">End-to-End</a> deep learning. We will be focusing only on the steering angle in this project, and the network will use each of the simulator’s frames as input.</p><h3>Deep Learning</h3><p>If you want an introduction to Deep Learning and some useful resources about it, please read my previous <a href="https://hackernoon.com/traffic-signs-classification-with-deep-learning-b0cb03e23efb#.tmv9fduho">post</a>.</p><h3>Behavioral cloning Project</h3><p>Udacity provided us with a simulator to drive around a track and gather data for the model.</p><p>Here is a preview:</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FsKqI5fNMQ4o%3Ffeature%3Doembed&amp;url=http%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DsKqI5fNMQ4o&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FsKqI5fNMQ4o%2Fhqdefault.jpg&amp;key=d04bfffea46d4aeda930ec88cc64b87c&amp;type=text%2Fhtml&amp;schema=youtube" width="640" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/ecbfb82362efc8e8d5d20bae508e31a4/href">https://medium.com/media/ecbfb82362efc8e8d5d20bae508e31a4/href</a></iframe><p>As you can see, there are 2 tracks available, plus an autonomous and a training mode. The goal is to drive around the tracks in the manual mode to “teach” our network how to steer the angle in different situations.</p><p>An exemple of data gathered while driving:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*grM8qsgebf1QDNyak1IDKg.png" /></figure><p>Each row contains the location of 3 images. Indeed, the car has 3 cameras installed. One in the center, one in the left, and one in the right. We also have data about the throttle, the steering angle, the break and the speed.</p><p>This project was the first one where I had full autonomy from start to end. From collecting the data, to computing the end result.</p><p>Here are some errors I made. I hope it will help you avoid them in the future.</p><p><strong>Always analyze your data</strong> and take a close look. Don’t just assume.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1012/1*3gGKq-8880bbU3C463jdfQ.png" /></figure><p>The steering distribution I had was biased towards the 0. I was using a keyboard and the steering angle went too fast to -+0.4</p><p>This is the original result I got:</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FhU3glJ3k9hs%3Ffeature%3Doembed&amp;url=http%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DhU3glJ3k9hs&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FhU3glJ3k9hs%2Fhqdefault.jpg&amp;key=d04bfffea46d4aeda930ec88cc64b87c&amp;type=text%2Fhtml&amp;schema=youtube" width="640" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/0c400cce81463699e0fa7d777f4aaf9a/href">https://medium.com/media/0c400cce81463699e0fa7d777f4aaf9a/href</a></iframe><p>If you notice, the angles were going to +-25° a lot. It actually took me a while to notice that. Instead of debugging the actual value my network was giving I was coming up with non related solutions:</p><ul><li>The data recorded is not good.</li><li>The network is not good.</li><li>The data augmentation is not good.</li></ul><p>After some time, I realized I simply forgot to normalize the RGB values of the simulator’s frame during the autonomous drive. The values being between 0 and 255 instead of 0 and 1, the network would output steering angles that were way too high</p><p><strong>Always remember to use the same image processing when you build your network and when you actually use it!</strong></p><p>After I used a joystick, drove slowly to minimize the steering angle at 0, used the left and right images, and randomly flipped images, I got this data:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1014/1*51mU7YPVrmcF5QXq51A7dw.png" /></figure><p>The model was still not good enough. It seems that it was overfitting to the track and would sometimes turn too soon. I used more data augmentation to avoid this. I added some random brightness in my images.<br>Also, instead of using every image I have, I started to randomize the selection.</p><p>This seemed to help avoid overfitting and the car made a whole lap without falling!</p><p><strong>Understand overfitting. More data might not be the answer.</strong></p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FbQEFQ4RdE84%3Ffeature%3Doembed&amp;url=http%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DbQEFQ4RdE84&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FbQEFQ4RdE84%2Fhqdefault.jpg&amp;key=d04bfffea46d4aeda930ec88cc64b87c&amp;type=text%2Fhtml&amp;schema=youtube" width="640" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/1b52f3a596cf071d1c38b89ff19f8e61/href">https://medium.com/media/1b52f3a596cf071d1c38b89ff19f8e61/href</a></iframe><h3>Network architecture</h3><p>I buildt a small network with 6 CNNs, comprised of 32, 64 and 128 filters of size 3X3.</p><p>I put a maxpool after every 2 CNNs and one dropout after the first 4 ones. I wanted to avoid overfitting the model.</p><p>These convolution layers were followed by 3 fully connected layers. I also added two dropouts there. All the convolution blocks and the FC layers had exponential relu (ELU) as activation function. This in my opinion would help generate smoother angles.</p><p>I trained the model using the keras generator with a batch size of 256 for 6 epochs. In each epoch, I generated ~35000 images randomly. This was done to help the model generalize and not overfit.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/302/1*QSrTbyFWSwz_udUk72oAqQ.png" /></figure><h3>Machine learning is hard</h3><p>I have been a software engineer for a decade now, and both fields have a lot of similarities:</p><p>Contrary to common belief, the difficulties are not due to math. There are many frameworks and awesome ressources that abstract the low level and help you build a deep learning system with a few lines of code.</p><p>The difficulty is to build an intuition to understand how and what tools to use, as well as the available algorithms and trade-offs.</p><p>What makes Machine learning even harder is twofold.</p><ul><li>It is way harder to understand what is wrong. Debugging is still very hard. Many tools are lacking to efficiently understand where you should focus your energy. The data, the model, the preprocessing, the algorithm?</li><li>On top of that, the debugging cycle is long. It can take hours before you can see the result of a change you made.</li></ul><p>I can only imagine the possibilities with more powerful computers and better tools.</p><h3>To be continued…</h3><p>Next time, we’ll go back to computer vision techniques to <a href="https://medium.com/@MSqalli/lane-detection-446986c44021#.oha5x2uw7">enhance our Lane detection</a>.</p><p>Here is a preview:</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2F-Cb7HKFeJ5E%3Ffeature%3Doembed&amp;url=http%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3D-Cb7HKFeJ5E&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2F-Cb7HKFeJ5E%2Fhqdefault.jpg&amp;key=d04bfffea46d4aeda930ec88cc64b87c&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/c708c813ed0cd81eae5c259439b7aab4/href">https://medium.com/media/c708c813ed0cd81eae5c259439b7aab4/href</a></iframe><p><strong>Stay tuned!</strong></p><p>Also, please feel free to comment, ask questions, and give advice.</p><p><em>Thanks to </em><a href="https://medium.com/u/d8603e818394"><em>Eric Sauvegrain</em></a><em> for his feedback on this post.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b54e921f64d2" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Traffic signs classification with Deep Learning.]]></title>
            <link>https://medium.com/hackernoon/traffic-signs-classification-with-deep-learning-b0cb03e23efb?source=rss-151f42ef18b8------2</link>
            <guid isPermaLink="false">https://medium.com/p/b0cb03e23efb</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[self-driving-cars]]></category>
            <category><![CDATA[tech]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[deep-learning]]></category>
            <dc:creator><![CDATA[Mehdi Sqalli]]></dc:creator>
            <pubDate>Mon, 05 Dec 2016 19:13:13 GMT</pubDate>
            <atom:updated>2019-05-08T08:38:35.069Z</atom:updated>
            <content:encoded><![CDATA[<p><em>You can find my code </em><a href="https://github.com/MehdiSv/TrafficSignsRecognition/"><em>here</em></a><em>. It might be more up to date than the article ;).</em></p><h3>Intro</h3><p>This is the second article in the self driving cars series. If you want to know why I’m sharing this and more about my journey, please read <a href="https://medium.com/@MSqalli/self-driving-car-journey-to-build-one-5413ad0e4115">this</a>.</p><h3>Goal</h3><p>Classify traffic signs using a simple convolutional neural network.</p><h3>Machine Learning</h3><p>Imagine you need to build a program that recognizes written digits.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/218/0*lYOVf4meXn0tOCik." /><figcaption>Image from the MNIST database.</figcaption></figure><p>This is a 5. But it could arguably be a 3.</p><p>What cutoffs/rules, would you use to go from a 3 to a 5?</p><p>Instead of trying to handpick all the rules and build a very complicated program, researchers have decided to show a computer thousands of examples and let it try to solve the problem by experience. This is the beginning of machine <a href="https://hackernoon.com/tagged/learning">Learning</a>.</p><p>One of the main problems with machine learning is feature extraction. Even though we show the computer thousands of examples, we still needed to tell him what features it should focus on. For complex problems, this was not good enough.</p><p>Deep learning models circumvent that. They learn by themselves what features they should focus on.</p><h3>Deep Learning</h3><p>For the sake of brevity, I’m not going to dive in the mathematical explanations of how deep learning works. It took me around 20 hours to understand the concepts and use them. Instead, I’ll try to explain the intuition behind deep learning. I’ll post some videos and lectures I used if you want to go deeper. No pun intended.</p><p>As humans, recognizing an object seems like a pretty simple task. There is hardly any effort involved on our part, at least not consciously. But there is actually a lot of work done by our brain before we can really understand what we’re looking at.</p><p>In the late 1950s, David Hubel and Torsten Wiesel, two famous neurophysiologists, made experiments on a cat to show how the neurons in the visual cortex work.</p><p>For one, they showed that nearby cells process information from nearby visual fields, forming a topographical map. Moreover, their work determined that neurons with similar functions are organized into columns, tiny computational machines that relay information to a higher region of the brain, where a visual image is progressively formed.</p><p>The brain basically combines low level features such as basic shapes, curves and builds more complex shapes out of it.</p><p>A deep learning convolutional neural <a href="https://hackernoon.com/tagged/network">network</a> is similar. It first identifies low level features and then learns to recognize and combines these features to learn more complicated patterns. These different levels of features come from different layers of the network.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/717/0*FTEdYfedHSLtBkXp." /></figure><p>Deep Learning is a fascinating field and I hope I gave you a clear enough introduction. I encourage you to watch the <a href="https://www.youtube.com/playlist?list=PLLvH2FwAQhnpj1WEB-jHmPuUeQ8mX-XXG">wonderful Stanford class about the subject</a>.</p><p>If you prefer reading, I’d advise you Goodfellow, Bengio, and Courville’s <a href="http://www.deeplearningbook.org/">book</a>.</p><h3>Traffic signs project</h3><p>Detecting and Classifying Traffic signs is a mandatory problem to solve if we want self driving cars.</p><p>The dataset we will be using is a <a href="http://benchmark.ini.rub.de/?section=gtsrb&amp;subsection=dataset">German Traffic sign dataset available online</a>.</p><p>It contains more than 50,000 images in total, divided into 43 different classes: speed limits, dangerous curves, slippery road… Here are some of them.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/353/0*cr7R5YbCwLsDUHDj." /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/353/0*5TH7sjRjihK5IQUq." /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/353/0*Fnu7u3GONyHqQzG0." /></figure><p>This dataset was used in a competition a few years ago. The best result for the competition correctly guessed 99.46% of the signs. In comparison, human performance was established at 98.84%. Yes, the machine was more efficient than the human, as it was better at handling the most difficult cases, such as a blurry image of a speed limit sign that could be mistaken for a different speed limit.</p><h3>Data augmentation</h3><p>Before we start building our deep learning network, let’s analyze the data.</p><p>Here is the distribution of the different classes:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/555/0*q57IENEOvI09SUeu." /></figure><p>As you can see there is too much difference between the classes. We are going to create some data to balance the number of inputs and reduce the probable bias the network could have towards some classes. It will also help us give more data to our network.</p><p>An easy way to do that is take the images and rotate them by a few degrees. By adding 5, -5, 10, and -10 degrees, we can already increase the input of some classes fivefold.</p><p>This is the distribution we get after this operation.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/555/0*wxU_pLI9sOAPzj4L." /></figure><p>The data is more balanced, and each class has at least 500 images.</p><h3>Network architecture</h3><p><em>This part of the article describes in detail the network architecture. If you are not familiar with deep learning, you can skip to the results section.</em></p><p>First layer is a CNN with a patch size of 3*3, a stride of 1, SAME padding and a depth of 64.</p><p>Second and third layers are fully connected layers with a width of 512.</p><p>The final layer is a fully connected layer with a width of 43 (the amount of classes).</p><p>I used the Adam optimizer with its default parameters as it is currently regarded as the most efficient.</p><p>I used a batch size of 250 and 100 training epochs.</p><p>I tried adding more convolutional networks but they didn’t improve the results and increased the computation time by a lot. It didn’t feel necessary to add them as there is a low statistical invariance between the pictures we work on. Most of them are already centered and cropped around the sign.</p><p>I used a medium sized network as the signs are overall pretty simple in shape and color, but wide fully connected layers as there are some variations in the sign’s shapes, colors and overall appearance.</p><p>I didn’t use more than 100 epochs as the accuracy wasn’t improving after that. I also wanted to keep the network light so training it wouldn’t take too much time on AWS and the results are satisfying as is.</p><h3>Results</h3><p>Up until this point I only authorized myself to look at the validation accuracy. When I was above 99% and happy with the result, I decided to compute against the testing set.</p><p><em>Final validation accuracy: 0.9938</em></p><p><em>Final test accuracy: 0.9110</em></p><p>The 8% drop between the validation and test sets shows that I overfit my model.</p><p>Here are some things I could try to avoid that:</p><ul><li>L2 Regularization.</li><li>Early stopping.</li><li>Dropout.</li></ul><p>A look at the confusion matrix also helps us see where we had most errors:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1003/0*WpK8lP-7XjHlfovC." /></figure><p>As you can see, the majority of errors are located on the top left. These are all the speed limits. It seems like the model can detect it’s a speed limit but has trouble identifying the difference between the values in the sign.</p><h3>Extra</h3><p>I tested some random signs downloaded from the internet to see how the model would react:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/32/0*qVLQcBQZvtWIS0Vz." /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/32/0*jrXB3lND4KOnLH_M." /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/32/0*nyiVKKcGqClZ6JR4." /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/32/0*N6YcXijgfsE3GTlt." /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/32/0*hZMsssbtqT3rlrjG." /></figure><p>The two last signs were not in our dataset.</p><p>The three first signs were recognized correctly with a confidence of 99+%</p><p>The 4th sign was (obviously) wrongly recognized with a confidence of 80% as a No entry sign.</p><p>The 5th sign was (obviously) wrongly recognized with a confidence of 99% as a speed limit. The model should do better than that.</p><h3>Thoughts</h3><p>It took me around 20 hours to go from 0 knowledge in deep learning to being able to implement a simple small network. A few hours later I was able to get satisfying results.</p><p>Deep learning is really impressive. I can’t wait to see how much I can still learn and progress in this domain.</p><p>Next time we’ll talk about behavioral cloning. We’ll drive a car in a simulator, and see how well our deep learning model can learn from that to drive by itself. Sounds exciting right?! <strong>Stay tuned!</strong></p><p>Also, please feel free to comment, ask questions, and give advice.</p><p><em>Thanks to </em><a href="https://medium.com/u/d8603e818394"><em>Eric</em></a><em>, </em><a href="https://medium.com/u/e027d96f8057">Aboud</a><em>, </em><a href="https://medium.com/u/1574406fd15">hamri said</a>, and <a href="https://medium.com/u/288af6ec97d0">Hamza</a> <em>for their feedback on this post.</em></p><figure><a href="http://bit.ly/HackernoonFB"><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*0hqOaABQ7XGPT-OYNgiUBg.png" /></a></figure><figure><a href="https://goo.gl/k7XYbx"><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Vgw1jkA6hgnvwzTsfMlnpg.png" /></a></figure><figure><a href="https://goo.gl/4ofytp"><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*gKBpq1ruUi0FVK2UM_I4tQ.png" /></a></figure><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fupscri.be%2Fdde502%3Fas_embed%3Dtrue&amp;dntp=1&amp;url=https%3A%2F%2Fupscri.be%2Fhackernoon%2F&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=upscri" width="800" height="400" frameborder="0" scrolling="no"><a href="https://medium.com/media/3c851dac986ab6dbb2d1aaa91205a8eb/href">https://medium.com/media/3c851dac986ab6dbb2d1aaa91205a8eb/href</a></iframe><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b0cb03e23efb" width="1" height="1" alt=""><hr><p><a href="https://medium.com/hackernoon/traffic-signs-classification-with-deep-learning-b0cb03e23efb">Traffic signs classification with Deep Learning.</a> was originally published in <a href="https://medium.com/hackernoon">HackerNoon.com</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Lane detection]]></title>
            <link>https://msqalli.medium.com/lane-detection-446986c44021?source=rss-151f42ef18b8------2</link>
            <guid isPermaLink="false">https://medium.com/p/446986c44021</guid>
            <category><![CDATA[udacity]]></category>
            <category><![CDATA[learning]]></category>
            <category><![CDATA[computer-vision]]></category>
            <category><![CDATA[tech]]></category>
            <category><![CDATA[self-driving-cars]]></category>
            <dc:creator><![CDATA[Mehdi Sqalli]]></dc:creator>
            <pubDate>Sun, 06 Nov 2016 18:49:26 GMT</pubDate>
            <atom:updated>2016-11-07T02:33:58.405Z</atom:updated>
            <content:encoded><![CDATA[<p><em>You can find my code</em><a href="https://github.com/MehdiSv/FindLanes"><em> here</em></a><em>. It might be more up to date than the article ;).</em></p><h3>Intro</h3><p>This is the first article in the self driving cars series. If you want to know why I’m sharing this and more about my journey, please read <a href="https://medium.com/@MSqalli/self-driving-car-journey-to-build-one-5413ad0e4115">this</a>.</p><h3>Goal</h3><p>For each image of the given video, we are going to detect the driving lane lines on each side of the car and visually highlight them.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*LtAD-SDzJWGFL4BK." /></figure><h3>Input conditions</h3><p>The first iteration of the lane detection algorithm will work well on non-curved roads with a single camera on the dashboard. In these images, the weather conditions are good, and it is daytime. We will handle shadows, different line colors and continuous/discontinuous lanes. Much more work will need to be done later to make this work on curved roads, in different lighting and weather conditions.</p><h3>Algorithm: Step by step</h3><p>This algorithm is executed step by step on each frame:</p><p>1: Convert the picture to HSV. If you’re unfamiliar with HSV, it is basically a different way to describe color than the classic RGB we are used to, you can learn more about it on <a href="https://en.wikipedia.org/wiki/HSL_and_HSV">Wikipedia</a>. We will explain later why this conversion is necessary.</p><p>2: We apply a small gaussian blur on the picture. It helps us reduce noise, keeping us from detecting tiny, irrelevant details on the picture such as foliage or very distant objects.</p><p>3: We now want to filter out (turn black) any pixel that is not the color of the lanes, in this case white or yellow.</p><p>This is where the HSV conversion comes in handy: it allows us to filter these colors much more easily by focusing on the hue and saturation of the pixel (which color it is and how intense the color is), and not so much on its value (how dark it is), thus handling shadows and overall worse lighting conditions much more easily.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/800/0*XKjgDkMrOykvtFm8." /></figure><p>The image has way fewer elements now, there only remain ones with white or yellowish tint, with some false positives but our two lines are very clearly defined.</p><p>4: We use Canny edge detection. The Canny algorithm detects edges on a picture by looking for quick changes in color between a pixel and its neighbors, in our case between the white or yellow lane line and black pixels.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/800/0*i_0roJoCSyFi4vO2." /></figure><p>5: We define our region of interest now. Given the position and orientation of the camera, we know that the lanes will be located in the lower half of the image, usually in a trapezoid covering the bottom corners and the center. We don’t want our region to be too narrow and have the lines out of our region of interest.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/800/0*zBdioxMJFD_qWmW5." /></figure><p>6: We run Hough transform to detect lines on our image. What does Hough transform do? To summarize quickly, it allows us to easily extract all the lines passing through each of our edge points and group by similarity, thus grouping together the points that are roughly located on a same line. This algorithm is overall fascinating and I encourage you to take a look at it here:<a href="https://classroom.udacity.com/courses/ud810/lessons/2870588566/concepts/34824487840923#"> Intro to Computer Vision @ Udacity</a>. You need to create an account on<a href="https://medium.com/u/2929690a28fb"> Udacity</a> but it’s worth it, believe me.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/800/0*pK7R_qThbLpcpaDs." /></figure><p>The Hough transformation gives us back multiple lines, but we only want 2 distinct lanes for our car to drive in between. There is still some work to do.</p><p>7: We filter the lines with horizontal slopes first, to make sure the data left is clean.</p><p>8: We distinguish between the left and right lane thanks to the line slopes.</p><p>9. Now that we grouped our points by lane line, we compute the linear regression of both groups and draw the results in our region of interest. Simply put, a linear regression is an attempt at finding the relationship between a group of points, so basically finding the line that passes at the closest possible distance from each point. This operation allows us to fill the blanks in a dashed lane line.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/800/0*dhWAr_pvFJ_0XUlt." /></figure><p>Beautiful.</p><p>10: All we have to do now is add these lanes to our original image.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/800/0*qEjm9xSPQPk0_f7X." /></figure><p>As you can see, since the lanes are curved, and our extrapolation being linear, we do not match the curbs perfectly. But this is just an introduction and we’ll consider it good enough for now ;).</p><p>11: The final tweak we add since we will be using our algorithm on videos is some smoothing. We always take into consideration the previous lanes by giving them a weight before drawing the new lanes to have a smoother movement and reducing the effect of errors.</p><p><strong>Here are 3 videos to show the final output.</strong></p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FO9ER85dBeoQ%3Ffeature%3Doembed&amp;url=http%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DO9ER85dBeoQ&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FO9ER85dBeoQ%2Fhqdefault.jpg&amp;key=d04bfffea46d4aeda930ec88cc64b87c&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/3295f2e2231878197ed9991da726483b/href">https://medium.com/media/3295f2e2231878197ed9991da726483b/href</a></iframe><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FSHfWJpK02ng%3Ffeature%3Doembed&amp;url=http%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DSHfWJpK02ng&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FSHfWJpK02ng%2Fhqdefault.jpg&amp;key=d04bfffea46d4aeda930ec88cc64b87c&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/6f4c4f0948c4e0602d4825f01a6cc1fb/href">https://medium.com/media/6f4c4f0948c4e0602d4825f01a6cc1fb/href</a></iframe><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2F2A33l-t1Fe8%3Ffeature%3Doembed&amp;url=http%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3D2A33l-t1Fe8&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2F2A33l-t1Fe8%2Fhqdefault.jpg&amp;key=d04bfffea46d4aeda930ec88cc64b87c&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/b4576a81b45cc6c69a4d483966f814ef/href">https://medium.com/media/b4576a81b45cc6c69a4d483966f814ef/href</a></iframe><p>Next time we’ll talk about deep learning and how we can detect traffic signs! <strong>Stay tuned.</strong></p><p><em>Thanks to </em><a href="https://medium.com/u/d8603e818394"><em>Eric Sauvegrain</em></a><em> for his feedback on this post.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=446986c44021" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Self-driving car: Journey to build one]]></title>
            <link>https://medium.com/spot-angels/self-driving-car-journey-to-build-one-5413ad0e4115?source=rss-151f42ef18b8------2</link>
            <guid isPermaLink="false">https://medium.com/p/5413ad0e4115</guid>
            <category><![CDATA[parking]]></category>
            <category><![CDATA[tech]]></category>
            <category><![CDATA[self-driving-cars]]></category>
            <category><![CDATA[learning]]></category>
            <category><![CDATA[self-improvement]]></category>
            <dc:creator><![CDATA[Mehdi Sqalli]]></dc:creator>
            <pubDate>Sun, 06 Nov 2016 18:46:28 GMT</pubDate>
            <atom:updated>2016-12-05T19:30:13.492Z</atom:updated>
            <content:encoded><![CDATA[<p>I am the Co-Founder and CTO @<a href="https://medium.com/u/9738038d1aa9">Spot Angels</a>. Our goal is to fix parking for today and tomorrow’s cars by building real-time parking maps (If you want to know more about it, please read the <a href="https://medium.com/spot-angels/should-parking-be-fixed-4c4fd8e4861c">article</a> my Co-Founder <a href="https://medium.com/u/288af6ec97d0">Hamza</a> wrote).</p><p>I’ve always been fascinated by computer vision and its application. I worked on a Motion &amp; Tracking software at my first startup 4 years ago. At Spot Angels, we use computer vision to build part of our maps. I am personally excited by the potential applications of all the imagery data we will collect from cars as they get more and more sensors.</p><p>Self-driving cars were the best way to keep learning and understand the current trend. I decided to join the first cohort of the self-driving car nanodegree @<a href="https://medium.com/u/2929690a28fb">Udacity</a>. I intend to follow this 9 month program during my free time. My goal is to learn how to build a self-driving car at the end of the program.</p><p>I wanted to share what I learn with our team. Everybody is curious to understand the technology behind self driving cars. I thought others might benefit from it so I decided to share this publicly.</p><p>List of publications so far:</p><ul><li><a href="https://medium.com/@MSqalli/lane-detection-446986c44021#.y2d6ub2j0">Lane detection</a>.</li><li><a href="https://medium.com/@MSqalli/traffic-signs-classification-with-deep-learning-b0cb03e23efb#.jeuf553hj">Traffic signs classification with Deep Learning</a>.</li></ul><p>I will be updating this post to list all the projects we worked on.</p><p><strong>Stay tuned.</strong></p><p><em>Thanks to </em><a href="https://medium.com/u/288af6ec97d0"><em>Hamza</em></a><em> for his feedback on this post.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=5413ad0e4115" width="1" height="1" alt=""><hr><p><a href="https://medium.com/spot-angels/self-driving-car-journey-to-build-one-5413ad0e4115">Self-driving car: Journey to build one</a> was originally published in <a href="https://medium.com/spot-angels">SpotAngels</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>