<?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 Alexabriannaguzman on Medium]]></title>
        <description><![CDATA[Stories by Alexabriannaguzman on Medium]]></description>
        <link>https://medium.com/@alexabriannaguzman?source=rss-95a5682012c5------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*tiP7dvu5UYafhDiD</url>
            <title>Stories by Alexabriannaguzman on Medium</title>
            <link>https://medium.com/@alexabriannaguzman?source=rss-95a5682012c5------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Tue, 17 Nov 2026 23:05:44 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@alexabriannaguzman/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 a drone fly fast through unknown environments using only a camera, and what’s still stopping it…]]></title>
            <link>https://medium.com/@alexabriannaguzman/can-a-drone-fly-fast-through-unknown-environments-using-only-a-camera-and-whats-still-stopping-it-6f2de855f516?source=rss-95a5682012c5------2</link>
            <guid isPermaLink="false">https://medium.com/p/6f2de855f516</guid>
            <category><![CDATA[lidar]]></category>
            <category><![CDATA[autonomous-driving]]></category>
            <category><![CDATA[drones]]></category>
            <category><![CDATA[navigation]]></category>
            <dc:creator><![CDATA[Alexabriannaguzman]]></dc:creator>
            <pubDate>Mon, 26 Oct 2026 06:54:44 GMT</pubDate>
            <atom:updated>2026-10-26T06:54:44.156Z</atom:updated>
            <content:encoded><![CDATA[<h3>Can a drone fly fast through unknown environments using only a camera, and what’s still stopping it from replacing LiDAR?</h3><p>Picture a drone flying through a compact forest at 50km/hr. No map, no GPS guiding it around the trees, no pilot either. Just cameras and a computer making decisions faster than you can blink. It has to spot every branch, every aspect of wildlife, not to mention it has to swerve over and over again without crashing.</p><p>To do this, some drones carry LiDAR, primarily for obstacle avoidance and indoor stability, making them incredibly easy to fly. Unlike standard cameras that only capture surface textures, LiDAR (Light Detection and Ranging) is an active remote sensing technology that uses pulsed laser beams to measure distance and create precise, three-dimensional maps of objects and the environment. It works, but it’s considered to be problematic; it’s heavy, highly expensive, and has weather and interference limitations. Meanwhile, we see lightweight cameras in our everyday lives; your phone already has a few. So why can’t a drone fly fast through unknown environments using only a camera, and what’s stopping it from replacing LiDAR?</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Txe8iG9ESlYyzRUaYcNbcQ.png" /><figcaption>LiDAR uses laser measurements to build a three-dimensional representation of the surroundings</figcaption></figure><h4><strong>What does a drone actually need to know to fly somewhere it has never been before?</strong></h4><p>When a drone is flying through an environment it has never encountered before, it cannot simply follow a pre-programmed route. Instead, it must make decisions based on what is happening around it. This becomes especially difficult when the drone is moving at a higher speed, as the surrounding environment constantly changes, altering its perspective.</p><p>But seeing the environment is not the same as understanding it.</p><p>A camera gives the drone a variety of images, but an image by itself doesn’t tell the drone exactly what it should do. The drone needs to extract useful information from them. It has to determine which parts of the environment are obstacles, which areas are open, how far away objects are, and how its own movement is changing its view of the environment. All of this has to happen while the drone is already flying.</p><p>This becomes even more complicated because the drone’s view is constantly changing, adjusting its view every passing moment. Imagine walking through a forest while looking through a camera. Trees that were far away a moment ago appear much closer now, and some even disappear in the distance. Even though the trees are not moving, they appear to move in the camera’s view because you are moving. A drone cannot simply analyze one image and decide where to go, it must continuously update its understanding of the environment as new images arrive. It needs to connect what it sees to what it saw moments ago, to understand how it’s moving in the area.</p><p>Now imagine doing all this at high speed.</p><p>If a drone is moving slowly, it has more time to process what the camera sees and change its path. However, when a drone is moving quickly, the time it takes to analyze obstacles and its trajectory becomes much smaller.</p><p>This scenario creates the high-speed obstacle avoidance problem.</p><p>The drone now has to make decisions in a very limited amount of time. It needs to detect an obstacle, measure the distance between the obstacle and itself, and determine whether the path it’s taking will bring it too close to an obstacle and if it needs to change its trajectory. If the drone cannot process this information and respond quickly enough, it may increase the risk of collision.</p><p>Autonomous navigation is not just about what it is seeing, but it’s also about turning what the camera sees into a prediction of what is going to happen if it keeps moving one way.</p><p>This brings me to a much more difficult question:</p><h4><strong>How can a camera, which produces flat 2D images, give a drone information about the 3D world around it?</strong></h4><p>A typical camera only captures two-dimensional images. It records what the world looks like from a specific perspective, but it does not directly tell the drone how far away each object is. For a person looking at a photo, depth is easy to estimate because our brains have learned to interpret visual clues. For a drone, however, this information has to be calculated.</p><p>The drone can use motion as a method to estimate depth. As a drone moves forward, nearby objects appear to shift position around the camera’s view much quicker than those far away. This is called <a href="https://thedecisionlab.com/reference-guide/psychology/motion-parallax">motion parallax</a>. The trees passing right by the drone might seem to be moving to the side of the image very quickly, while mountains in the distance seem to barely move position.</p><p>The drone can use these changes between a sequence of images to estimate how objects are positioned in three-dimensional space. This process is referred to as <a href="https://www.sciencedirect.com/topics/computer-science/visual-odometry">visual odometry</a>. Rather than using wheels or some other form of measurement, the drone can estimate its own movement by tracking how features in the camera images change over time.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/858/1*WVoHNoBhYco6EKwkS9V5qw.png" /><figcaption>By comparing features across successive images, a drone can estimate its movement relative to nearby objects</figcaption></figure><p>For example, imagine the drone sees a certain branch in one frame. After half a second, the drone captures an image of that same branch in a different position within the image. By comparing the two frames, the system can estimate how the drone has moved relative to the branch.</p><p>But there is a catch.</p><p>In the real world, environments are messy. A forest, for example, might be hundreds of tree branches that look similar. There could be shadows that change the appearance of objects, lighting that can suddenly become brighter or darker, and not to mention that parts of the environment can disappear from view.</p><p>There are also situations where the drone needs to differentiate between its own motion and another motion caused by something else. If an object appears to shift across the image, that could be because the object moved, or because both happened at once.</p><p>This becomes especially difficult when the drone is flying somewhere unfamiliar. It cannot rely on a perfectly prepared map, so it will have to build its understanding of the environment from the data it collects. The faster the drone flies, the more challenging it becomes. Objects can move more significantly across the camera’s field of view between frames, giving the system less time to process what it sees.</p><p>So the camera is doing much more than simply recording the environment. Its images provide the information the drone uses to estimate its movement, depth, and the structure of the world around it.</p><p>But even if a drone can build this 3D understanding, there is still another problem.</p><p>Knowing what is around you doesn’t automatically tell you where you should go.</p><p>The drone now has to turn everything it has perceived into a decision:</p><h4>Can one camera see enough of the world for a drone to fly safely?</h4><p>Even if a camera can give a drone a 3D representation of the drone’s surroundings, a basic limitation remains: it can only see what is visible from the perspective of its camera. Any other object outside that view is essentially invisible to the drone until it enters the frame.</p><p>This can be a major problem when flying through a complex environment. Imagine a drone moving between trees; it might have a clear view directly ahead, but a branch could be hidden behind another branch, or in the worst-case scenario, the drone might see it only at the last moment, after it has already gotten close.</p><p>The problem becomes even more important when the drone is moving fast. At 50km/hr, the drone travels almost 14 meters each second. That means something appearing at the edge of the camera’s view may become a concern within seconds. The drone has less time to react to anything it couldn’t see before.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*gB3fR81IvYRqeX5A_7a1Ow.png" /><figcaption>A single camera captures images within its field of view. Objects outside this area may be missed, creating blind spots that can make it harder for a drone to detect obstacles.</figcaption></figure><p>One way researchers can address this problem is by installing additional cameras on drones. Instead of relying on one viewpoint, cameras can be placed in different directions so the drone can get a better view of its surroundings. As a result, the drone will have several views of its surroundings and will be able to observe them more clearly. This can also reduce blind spots and give the system more information about what is happening around it.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*DUcv_wRo8rh8Au8Ngf3i8w.png" /><figcaption>Multiple cameras mounted around a drone capture views from different directions, helping it observe more of its surroundings and reduce blind spots.</figcaption></figure><p>Yet, adding additional cameras does not completely solve the problem. Each camera produces more information that the drone has to process. The system now has to combine these different views and interpret what they mean, all while the drone is in flight.</p><p>So the challenge is not only whether a camera can understand what it observes. It also needs to see enough of the environment in the first place. And once we start giving the drone more cameras to improve its vision, we create another problem: can a small drone process all of that information quickly enough?</p><h4><strong>How much computing power can a small drone actually carry?</strong></h4><p>The drone may be able to see its surroundings more clearly if it has more cameras, but processing all of that data is another issue. To make decisions while flying, the drone must quickly examine the images its camera constantly produces.</p><p>This requires plenty of processing power. To decide what to do next, the computer must interpret camera data and predict the drone’s movement. The amount of data to process increases significantly when multiple cameras are used.</p><p>However, weight matters, and a drone cannot simply carry an extremely powerful computer. A heavier computer requires more energy to keep the drone in the air, which can end up reducing its flight time. The drone therefore has to balance the weight and energy it can afford to carry.</p><p>There’s also the matter of speed. The faster the drone moves, the less time the computer has to process more information and make more complex decisions, but adding more hardware can make the drone heavier and less efficient. Researchers therefore need to make algorithms and hardware faster and more efficient without simply making the computer bigger.</p><p>The real issue with camera-only autonomous flight is that it’s not only a camera problem, but the entire system must keep pace with the increasing amount of information generated. If normal cameras and computers struggle to keep up with rapid movement, researchers have another question: could a completely different type of camera make high-speed flight easier?</p><h4><strong>Could a completely different type of camera make high-speed flight easier?</strong></h4><p>When it comes to equipping autonomous drones with vision systems, researchers aren’t limiting themselves to standard cameras. Various camera technologies can collect visual information in different ways, and some are built to handle the specific challenges that become important during fast movement.</p><p>One popular option is a stereo camera, which has two cameras positioned slightly apart from each other. By comparing the two perspectives, the drone can estimate depth and determine how far away objects are from each other. Stereo vision is already being used in small autonomous drones for real-time obstacle avoidance systems designed to work with limited weight and computing power.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/762/1*UPtiKO98EpJXXCgMkrwSvg.png" /><figcaption>Comparing depth maps from different camera systems. These images show how cameras can estimate the distance to objects, helping a drone understand the 3D layout of its surroundings.</figcaption></figure><p>Another option is an event camera. Rather than capturing complete images at regular intervals, event cameras detect changes in brightness at individual pixels and record exactly when those changes happen. This gives it very high temporal resolution, so it can respond to changes much faster than a traditional frame-based camera. This makes event cameras especially interesting for high-speed movement, where even small delays have significant consequences.</p><p>However, event cameras do have their own limitations. Since they mainly record changes rather than complete images, they can provide less information when a scene is mostly still. Researchers have also found challenges with processing event data and getting the same level of accuracy as conventional image-based systems.</p><p>This means there isn’t one camera that is simply better at everything. Different cameras have different strengths:</p><ol><li><strong>Monocular camera:</strong></li></ol><p>Advantages: Lightweight and inexpensive</p><p>Disadvantages: Cannot directly measure depth from a single image</p><p><strong>2. Stereo camera:</strong></p><p>Advantages: Can estimate depth using two camera views</p><p>Disadvantages: Requires two cameras and additional processing</p><p><strong>3. RGB-D camera:</strong></p><p>Advantages: Captures colour images and depth information</p><p>Disadvantages: Limited range; some types are mainly suited to indoor use</p><p><strong>4. Fisheye camera:</strong></p><p>Advantages: Wide field of view; useful for detecting obstacles in complex environmental</p><p>Disadvantages: Images can have noticeable lens distortion</p><p><strong>5. Event camera:</strong></p><p>Advantages: Respond quickly to changes in brightness; reduces motion blur</p><p>Disadvantages: Produces data that can be more difficult to process</p><p>Research is increasingly looking at combining different types of visual information rather than expecting one camera to solve every problem. For example, studies have explored combining conventional images with event-based data to reduce response time while keeping the detailed information provided by normal cameras. This suggests that the best solution might not be finding one perfect camera, but finding a way to combine different technologies so they can work together. However, even with these improvements, cameras still have limitations when it comes to measuring distances and detecting obstacles reliably in unfamiliar environments. This raises an important question: if cameras continue to improve, why haven’t they completely replaced LIDAR yet?</p><h4><strong>If cameras continue to improve, why haven’t they completely replaced LIDAR yet?</strong></h4><p>The real question isn’t just whether a camera can help a drone see what’s around it anymore. Researchers have made significant progress in autonomous navigation, but getting a system to work well in one environment is different from making it reliable in unfamiliar conditions. A camera-based system might perform well in a controlled testing area, but its performance can change when it encounters conditions it has not been trained to handle. Think about flying through a forest, sudden changes in lighting or unexpected surroundings can make visual navigation less reliable, pretty quickly.</p><p>For high-speed drone flight, even a small error could affect its ability to avoid obstacles. LiDAR offers another way to sense the world around it. Since it uses laser pulses to measure distances, it can provide information that does not rely heavily on interpreting camera images. This can make it a useful source of information when visual estimates are off. However, LiDAR also has limitations, including cost, weight, and sensitivity to certain environmental conditions. Neither technology is perfect, which is why researchers continue to investigate how autonomous drones can navigate reliably using different combinations of sensors.</p><p>Ultimately, the challenge is not just to make cameras more advanced, but to make camera-only navigation dependable in environments the drone has never encountered before. So, rather than asking whether cameras can replace LiDAR today, another question becomes more interesting: how close are researchers to making camera-only flight reliable enough to work in the real world?</p><h4>Can AI teach a drone to fly into the unknown?</h4><p>In 2021, researchers at the University of Zurich demonstrated that a drone could navigate through new environments at speeds of up to 40km/hr using only its built-in cameras and processing power. The drone was tested in environments including forests, buildings, ruins, and abandoned trains, where it had to find its way around obstacles without relying on a pre-existing map. This was an important step toward autonomous flight because the drone had to make decisions in places it had never encountered before. <a href="https://techxplore.com/news/2021-10-high-speed-drones-unknown-ai.html#google_vignette">(University of Zurich article)</a></p><p>The key to this achievement was how the researchers trained the drone. They first created a simulated environment where an expert algorithm could fly through obstacles and determine suitable flight paths. A neural network then learned from the expert’s decisions. Once trained, the network could use information from the drone’s onboard cameras to predict how it should fly in the real world. Instead of having to map the entire environment before planning a route, the drone could use what it saw to decide how to move.</p><p>Interestingly, the researchers found the simulation did not need to perfectly recreate the real world for the training to work. The system was able to transfer what it learned in simulation to real environments, even though they were not exact copies of the simulated ones.</p><p>However, this did not mean the problem was completely solved. Although the drone successfully flew through the tested environments at speeds of up to 40km/hr, the researchers identified faster sensors as one of the next steps toward safe flight at even higher speeds.</p><p>The experiment showed that camera-only navigation was possible at impressive speeds, but it also raised another question: even if a drone knows how to navigate an unfamiliar environment, can it react quickly enough when an obstacle suddenly approaches?</p><h4>Can a drone react fast enough to avoid a collision?</h4><p>While the University of Zurich’s 2021 research focused on navigating unfamiliar environments, another study from the same university explored a different challenge: reacting to obstacles quickly enough to avoid a collision. Published in 2020, the study investigated how event cameras could help drones respond to fast-moving objects.</p><p>In conventional camera systems, capturing an image, processing it, and deciding how to react can take 20–40 milliseconds. At high speeds, such delays can make a significant difference. To address this, researchers developed a system using event cameras, which detect changes in brightness across individual pixels rather than capturing complete images at regular intervals. Combined with a specially designed algorithm, this allowed the drone to detect approaching objects and calculate how to avoid them with an overall algorithm latency of just 3.5 milliseconds.</p><p>The researchers tested the system on an autonomous quadcopter using only onboard sensing and computing. They threw objects of different sizes and shapes toward the drone, both indoors and outdoors. In one particularly challenging test, the drone had to avoid a ball approaching from just three metres away at 10 metres per second. It successfully avoided the incoming objects more than 90% of the time in the reported experiments.</p><p>These results demonstrated how quickly event camera systems could respond to sudden changes in an environment. However, the study focuses on avoiding objects moving closer rather than navigating an entirely unknown environment. It therefore did not establish that event cameras alone could replace LiDAR for every type of autonomous flight.</p><p>Together, these two studies demonstrate different ways researchers are addressing the challenges of camera-based navigation. One shows how AI can learn to fly through unfamiliar surroundings, while the other shows how specialized cameras can help a drone react to approaching obstacles in milliseconds. Both represent important progress, but neither proves that camera-only flight is reliable in every possible situation.</p><h4>My opinion: Will cameras eventually replace LiDAR?</h4><p>Based on my research, I think the most promising solution is to combine different camera technologies rather than relying on just one. For example, I predict that future autonomous drones could combine stereo cameras, which detect changes extremely quickly. AI could then use the information from both cameras to understand the environment and decide how to navigate through it.</p><p>I predict that as these technologies improve, drones will become better at flying through unfamiliar environments at higher speeds without needing LiDAR. Combining cameras could help overcome some of the limitations of traditional cameras, while more advanced AI could make navigation more reliable. However, this would still require further research and testing to prove that the system could work safely in unpredictable real-world conditions.</p><p>Ultimately, I think camera-only drones could eventually replace LiDAR for certain purposes, especially where weight and cost matter. However, I don’t think LiDAR will become completely unnecessary. Instead, I predict that cameras will take over more navigation tasks as technology advances, while LiDAR will remain useful in situations where additional distance measurements are needed for reliability and safety.</p><h4>Citations:</h4><ol><li>Flying high-speed drones into the unknown AI: University of Zurich, 2021, <a href="https://techxplore.com/news/2021-10-high-speed-drones-unknown-ai.html#google_vignette">https://techxplore.com/news/2021-10-high-speed-drones-unknown-ai.html</a></li><li>Dynamic obstacle avoidance for quadrotors with event cameras: Falanga, Kleber &amp; Scaramuzza, 2020, <a href="https://www.science.org/doi/10.1126/scirobotics.aaz9712">https://www.science.org/doi/10.1126/scirobotics.aaz9712</a></li><li>Moving robot (evet camera dodgeball study): University of Zurich, 2020, <a href="https://www.news.uzh.ch/en/articles/2020/Moving_robot.html">https://www.news.uzh.ch/en/articles/2020/Moving_robot.html</a></li><li>How fast is too fast? The role of perception latency in high-speed sense and avoid: Falanga, Kim &amp; Scaramuzza, 2019, <a href="http://rpg.ifi.uzh.ch/docs/RAL19_Falanga.pdf">http://rpg.ifi.uzh.ch/docs/RAL19_Falanga.pdf</a></li><li>Visual odometry: Part 1, the first 30 years and fundamentals: Scaramuzza &amp; Fraundorfer, 2011, <a href="https://rpg.ifi.uzh.ch/visual_odometry_tutorial.html">https://rpg.ifi.uzh.ch/visual_odometry_tutorial.html</a></li><li>Vision-Based Navigation Techniques for Unmanned Aerial Vehicles: Review and Challenges, <a href="https://www.mdpi.com/2504-446x/7/2/89">https://www.mdpi.com/2504-446x/7/2/89</a></li><li>A Comprehensive Review of LiDAR Applications in Crop Management for Precision Agriculture, <a href="https://www.mdpi.com/1424-8220/24/16/5409">https://www.mdpi.com/1424-8220/24/16/5409</a></li><li>Motion parallax, <a href="https://thedecisionlab.com/reference-guide/psychology/motion-parallax">https://thedecisionlab.com/reference-guide/psychology/motion-parallax</a></li><li>Visual odometry, <a href="https://www.sciencedirect.com/topics/computer-science/visual-odometry">https://www.sciencedirect.com/topics/computer-science/visual-odometry</a></li><li>Event-based vision, <a href="https://portal.fis.tum.de/en/publications/event-based-vision-a-survey/">https://portal.fis.tum.de/en/publications/event-based-vision-a-survey/</a></li><li>Multimodal obstacle detection and collision avoidance for micro aerial vehicles, <a href="https://www.ais.uni-bonn.de/~holz/papers/ecmr_uav_2013.pdf">https://www.ais.uni-bonn.de/~holz/papers/ecmr_uav_2013.pdf</a></li><li>Computing limits and blind spots, <a href="https://pmc.ncbi.nlm.nih.gov/articles/PMC11126069/?utm_source=chatgpt.com">https://pmc.ncbi.nlm.nih.gov/articles/PMC11126069/?utm_source</a></li><li>Comparing technologies, <a href="https://www.mdpi.com/1424-8220/21/24/8253?utm_source=chatgpt.com">https://www.mdpi.com/1424-8220/21/24/8253?utm_source</a></li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=6f2de855f516" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>