"Brain" and "Eye": Analysis of Robot SLAM Navigation and Positioning Technology

LIDAR (LiDAR):

three-dimensional space

Working principle: By emitting a laser beam and measuring the time of flight (TOF) or phase difference (FMCW) of the reflected light, point cloud data with millimeter-level accuracy is generated.

mechanical lidar (such as Velodyne VLP-16) achieves 360 ° environmental coverage through rotating scanning, which is suitable for outdoor complex terrain mapping;

solid-state lidar (such as Livox Mid-40) is widely used in indoor service robots (such as AMR mobile robots) due to its mechanical structure design, which improves reliability and integration.

✅Typical scene: storage AGV uses it to scan the shelf roadway, even in the dark environment can accurately avoid obstacles, to achieve "blind not crash 」.

Vision Sensor:

Semantic Understanding and Lightweight Perception

✅Technical classification:

single/binocular camera: by image feature matching (such as identifying the corner, table corner) to determine the position, low cost but "eye is not good" (there is a scale fuzzy problem);

Depth Camera (RGB-D): It can take color photos and measure distance, which is equivalent to installing "3D glasses" on the robot. The sweeping robot uses it to distinguish the carpet from the floor, and will never run with slippers stupidly ~

✅AI bonus: Now the visual sensor has learned "deep learning", such as DeepVO model can be like humans "look at the road prediction", in the white wall corridor this "low texture hell" scene, positioning error than the traditional method reduced by 40%!

multi sensor fusion:

Cooperative Sensing

each individual sensor has a short board? Then let them team up!

data-level fusion: the laser radar 3D point cloud "paste" to the camera image, to generate a 3D map with color, the robot can recognize at a glance "moving is a person, not moving is a table 」;

decision level fusion: multiple sensors after their own judgment "voting decision", such as laser radar said "ahead of the obstacle", the camera said "is a static shelf", the robot will know "no detour 」 ~

front-end processing:

calculate where you are born in real time

✅Laser SLAM algorithm:

LeGO-LOAM: First separate the ground point from the wall point, and then quickly match the features. The calculation speed is 35% faster than the traditional algorithm. The industrial inspection robot walks around the 200-meter corridor with a positioning error of less than 20cm;

Cartographer: Google open source "map master", can be different angles of scanning data into a complete map, millet sweeping robot with it to draw your home map, even the bottom of the sofa will not leak sweep ~

✅Visual SLAM algorithm:

ORB-SLAM3: support for a variety of cameras, but also to move while correcting the map, the shopping mall Pepper robot in the crowd, all rely on it to maintain "not dizzy 」;

DSO direct method: skip the traditional "find feature points" step, directly analyze the pixel grayscale changes, in the solid color wall scene can also quickly locate, micro-UAV high-speed flight rely on its "eyes 」 ~

back-end optimization:

building global maps and correcting errors

✅Factor graph optimization: connect the position of the robot and the map points into a "network graph", and use mathematical methods to continuously optimize, just like adjusting the position of each piece when playing a puzzle, and finally put together a perfect map;

✅Dynamic object processing: encounter pedestrians, pets these "moving trouble", through the "inter-frame contrast" or "semantic segmentation" to identify dynamic objects, automatically "ignore" them, to avoid the map "ghost obstacles 」 ~

Lightweight Optimization:

makes the brain run on the "low-profile computer"

robots commonly used raspberry pie, microcontroller computing power is limited? Engineers have a coup:

model compression: with "knowledge distillation" to the complex model "weight loss", such as the ResNet-50 compression into MobileNetV3, the amount of calculation is reduced by 60%, the calculation force requirements drop;

hardware acceleration: FPGA chip is used to accelerate feature extraction, the speed is doubled, UAV obstacle avoidance delay is reduced from 30ms to 15ms, and "eye disease and brain fast" is truly realized 」 ~