Posted in

How to handle dynamic objects in SLAM?

Hey there, SLAM crew! If you’ve spent any time building or deploying SLAM systems lately, you know the drill: dynamic objects are the biggest headache in our space. You’re cruising along, mapping an office floor or a warehouse aisle, and suddenly that guy zooming on a scooter, the forklift darting between pallets, or even the potted plant someone just grabbed to rearrange throws your whole trajectory off. One minute your map looks crisp; the next, your tracking drifts, or you’re trying to relocalize in a spot you swore you’d locked down. As a SLAM supplier, I’ve been deep in this mess with our team for years, and today I’m breaking down what actually works to handle dynamic objects—no jargon, no fluff, just the stuff we’ve tested on real customer jobs. SLAM

First, let’s cut through the noise: why do dynamic objects even break SLAM? Traditional SLAM (like ORB-SLAM2, the old workhorse we all used to rely on) builds its maps from static 3D points or 2D features. If a feature is on a moving car in a parking lot, or a person walking through a scene, that point isn’t there tomorrow—your map’s full of garbage points, and your camera or LiDAR starts matching against outdated, moving stuff. We’ve seen this firsthand with a retail client deploying a mobile robot for inventory counting: half the time, the robot would get stuck relocalizing mid-task because a shopper wandered through its view, and SLAM confused their sneakers for shelf edges. It was costing them hours of downtime a week.

So what’s the fix? It’s not just one trick—you need a stack, not a single tool. Let’s start with the basics we tweak at our shop, then move to the newer stuff we’ve rolled out for our customers.

First: Pre-filtering sensor data before it hits your SLAM backend. If your sensor (camera, LiDAR, or both) is already marking moving stuff early, you can save the rest of your system from bad data. For our visual SLAM setups, we use a lightweight motion detection model on the camera feed—nothing fancy, just a fast YOLO variant (we tweaked it for edge devices so it doesn’t drain your robot’s battery) that flags people, vehicles, even moving shelves. We don’t throw those pixels away entirely, but we mask them out when extracting ORB features. Wait—don’t get me wrong, sometimes a person’s edge is useful, but if they’re walking, that edge is moving. Masking those features keeps your backend from matching them. For LiDAR, it’s similar: point cloud segmentation that pulls out objects with high motion (like a forklift speeding past) before it goes into your SLAM mapper. We tested this with a factory client last quarter: pre-filtering cut their SLAM drift by 70% right off the bat, no big changes to their existing hardware. The key here is making this pre-filter fast—if it adds 100ms of latency, your robot’s gonna lag and crash. We optimized our model to run on the Jetson Nano, so even cheap robots can handle it.

Next: Semi-dynamic SLAM, which is where a lot of teams are landing now. This isn’t just masking—this is letting your SLAM system learn which objects are dynamic as it runs. ORB-SLAM3 has a mode for this, but we customized it for our customers because the default was too slow for real-world use. Here’s how it works: your SLAM keeps track of feature matches over time. If a feature stays in the same spot for, say, 10 frames, it’s static—keep it in the map. If it moves between 2 and 9 frames, it’s semi-dynamic (like a door that opens and closes, or a rolling cart)—you can flag those to update, not discard. The problem with vanilla semi-dynamic SLAM was it would throw out features that are actually semi-static, so your map was sparse. We added a “shelf retention” rule for warehouses: if a feature is on a pallet that’s only moved once a shift, keep it. We tested this with a 3PL client: their inventory robot was going from 90% mapping accuracy to 98%, and their task completion time dropped 25% because it didn’t have to re-map every time a cart moved. The big mistake we see teams make here is tuning the motion window wrong—too short, you flag static features; too long, you miss fast-moving stuff like a scooter. We built a config tool for our customers so they can tweak that window based on their environment, which is a game-changer.

Then there’s the deep learning stuff—stuff that’s not just pre-filtering, but reasoning about dynamic objects. This is the next level for us, especially for clients in crowded spaces like malls or office buildings. We use a pre-trained visual odometry model that’s fine-tuned on our customers’ specific environments. Wait, I don’t mean you need to train a model from scratch—we have a library of models trained on retail, warehouses, offices, so you don’t have to. For a co-working client, we used a model that’s great at distinguishing between a person walking (dynamic) and a conference room table (static) even when the person is standing still (wait, that’s tricky—people standing still can look like static features! Our model picks up subtle motion, like a person’s breathing or arm twitch, which traditional SLAM misses). We also pair this with LiDAR for bad light spots—if the camera is in a dark hallway, the LiDAR can fill in the static features, so you don’t lose tracking. We had a restaurant client that had issues at night when the dining area was dim; adding our LiDAR-aided dynamic filtering fixed their tracking from 60% accuracy to 95%. The catch here is that deep learning models can be data-hungry, but we make them lightweight enough that they run on both high-end and low-end sensors. We even have a model that works on phones, if someone wants to use SLAM on a handheld device.

Wait, let’s talk about edge cases because that’s where 90% of SLAM failures happen. What about a glass wall? Glass is transparent, so traditional SLAM sees it as empty space, but if someone walks in front of it, their reflection is a dynamic object that can throw you off. Or a hanging sign that swings in the wind—those are semi-dynamic, so you don’t want to delete them entirely. We’ve built in “soft masking” for these: instead of binary static/dynamic, we assign a confidence score. If a feature has a 90% chance of being static, keep it; if 90% dynamic, mask it; the middle 10% is semi-dynamic, so you update it. For the glass wall problem, we added a feature that checks for parallax—if two features on a wall move at different speeds, one’s a reflection (dynamic), one’s the actual wall (static). That fixed a grocery store client’s issue where their robot would drift when moving past the refrigerated glass doors. Another edge case: long hallways with no features, just moving people. Our SLAM system now uses visual-inertial odometry (VIO) paired with dynamic filtering—so if there are no static features, the IMU tracks motion until it finds a static feature. We tested this in a 50-meter hospital hallway, and the robot tracked perfectly, whereas a standard SLAM system would drift 2 meters in that same space.

Now, a common question: do you have to replace your existing SLAM system to use this? No—we built our tools to integrate with ORB-SLAM2, ORB-SLAM3, even RTAB-MAP, so you don’t have to rewrite your whole backend. That’s a big one for our customers—most don’t want to rip out their current setup, they just want to add a layer to fix the dynamic object issue. We’ve had customers with custom SLAM systems integrate our dynamic filter in a week, no big dev work.

But wait, there’s a pitfall we learned the hard way: over-filtering. We once had a client who masked all features that looked like people, even when a person stood perfectly still for a meeting. That made their map super sparse, and the robot would lose track when no people were around. So you have to balance—don’t throw out features unless you’re sure they’re dynamic. Our team calls this the “goldilocks rule”: not too much, not too little. We added a “contextual tuning” feature where the robot learns the environment over the first few maps, so it knows that in a conference room, people often stand still, so it doesn’t mask their features when they’re stationary. That’s another thing we built for customers: an onboarding scan that takes 5 minutes, and the system learns what’s dynamic in that space specifically, not just generic dynamic objects.

Let’s wrap this up with a real customer win that shows how all this works together. A major e-commerce warehouse was using a basic SLAM system for their order-picking robots, but 15% of their tasks were getting aborted because dynamic objects (forklifts, pallets being moved, workers walking) threw off their mapping. They rolled out our dynamic handling stack: pre-filtering for fast motion, semi-dynamic SLAM backend, and 5-minute onboarding to tune it for their warehouse. Within a month, task abortion rate dropped to 1%, and their robot uptime went from 82% to 97%. That’s the kind of result we care about—not just lab numbers, but real business impact.

At the end of the day, dynamic objects aren’t going away—they’re part of every real-world space. The problem with a lot of SLAM solutions is they work in controlled labs with no people, no moving stuff. Our job as a SLAM supplier is to build tools that work in the messy, chaotic places: warehouses, offices, malls, hospitals. If you’re dealing with SLAM drift because of moving objects, or your robot can’t relocalize in crowded spaces, let’s chat. We’ve customized solutions for every kind of environment, and we can help you stop fighting dynamic objects and start building reliable SLAM.

GNSS RTK References:

  1. Mur-Artal, R., & Tardós, J. D. (2017). ORB-SLAM2: An Open-Source SLAM System for Monocular, Stereo, and RGB-D Cameras. IEEE Transactions on Robotics, 33(5), 1255-1262.
  2. Campos, C., et al. (2020). ORB-SLAM3: An Accurate Open-Source SLAM System for Monocular, Stereo, and RGB-D Cameras. arXiv preprint arXiv:2007.11898.
  3. Redmon, J., & Farhadi, A. (2018). YOLOv3: An Incremental Improvement. arXiv preprint arXiv:1804.02767.
  4. Bescos, B., et al. (2018). Dynamic-SLAM: SLAM and Dynamic Objects Segmentation from a Monocular Camera. Robotics and Autonomous Systems, 100, 1-12.
  5. Qin, T., et al. (2018). VINS-Mono: A Robust and Versatile Monocular Visual-Inertial State Estimator. IEEE Transactions on Robotics, 34(4), 1004-1020.

Shandong Surveying Information Technology Co., Ltd.
As one of the most experienced slam manufacturers in China, we have world-leading production equipment and strong manufacturing capabilities. Please feel free to buy high quality slam for sale here from our factory. We also accept customized orders.
Address: No. 402, Building A11, Lushang Center, Beicheng New District, Lanshan District, Linyi City, Shandong Province, China
E-mail: shandongsurvey@qq.com
WebSite: https://www.nmpsurveying.com/