LiDAR Robot Navigation
LiDAR robots move using a combination of localization and mapping, as well as path planning. This article will explain these concepts and show how they work together using an easy example of the robot achieving a goal within a row of crop.
LiDAR sensors have low power demands allowing them to prolong a robot's battery life and decrease the raw data requirement for localization algorithms. This allows for more iterations of SLAM without overheating GPU.
LiDAR Sensors
The sensor is the core of a Lidar system. It emits laser pulses into the environment. The light waves bounce off objects around them in different angles, based on their composition. The sensor measures how long it takes each pulse to return and uses that data to calculate distances. The sensor is usually placed on a rotating platform allowing it to quickly scan the entire surrounding area at high speed (up to 10000 samples per second).
LiDAR sensors can be classified based on whether they're intended for applications in the air or on land. Airborne lidars are often attached to helicopters or unmanned aerial vehicles (UAV). Terrestrial LiDAR is usually mounted on a robotic platform that is stationary.
To accurately measure distances the sensor must be able to determine the exact location of the robot. This information is recorded by a combination inertial measurement unit (IMU), GPS and time-keeping electronic. LiDAR systems use sensors to calculate the precise location of the sensor in time and space, which is then used to create an image of 3D of the environment.
LiDAR scanners can also identify various types of surfaces which is especially beneficial when mapping environments with dense vegetation. For instance, if a pulse passes through a canopy of trees, it will typically register several returns. Typically, the first return is attributable to the top of the trees and the last one is attributed to the ground surface. If the sensor records each pulse as distinct, this is called discrete return LiDAR.
The Discrete Return scans can be used to analyze the structure of surfaces. For instance the forest may yield one or two 1st and 2nd return pulses, with the final big pulse representing bare ground. The ability to separate these returns and record them as a point cloud allows to create detailed terrain models.
Once a 3D model of the environment is created the robot will be able to use this data to navigate. This process involves localization and creating a path to reach a navigation "goal." It also involves dynamic obstacle detection. This is the process that identifies new obstacles not included in the original map and adjusts the path plan according to the new obstacles.
SLAM Algorithms
SLAM (simultaneous mapping and localization) is an algorithm that allows your robot to map its surroundings, and then determine its position in relation to that map. Engineers make use of this information for a variety of tasks, such as path planning and obstacle detection.
To allow SLAM to work it requires sensors (e.g. the laser or camera), and a computer running the right software to process the data. You'll also require an IMU to provide basic information about your position. The result is a system that can precisely track the position of your robot in an unknown environment.
The SLAM process is complex, and many different back-end solutions are available. No matter which one you select, a successful SLAM system requires a constant interaction between the range measurement device and the software that extracts the data and the vehicle or robot itself. This is a highly dynamic process that is prone to an unlimited amount of variation.
As the robot moves, it adds new scans to its map. The SLAM algorithm analyzes these scans against the previous ones using a process called scan matching. This helps to establish loop closures. The SLAM algorithm is updated with its robot's estimated trajectory when the loop has been closed discovered.
The fact that the surrounding can change in time is another issue that makes it more difficult for SLAM. If, for example, your robot is walking down an aisle that is empty at one point, and then comes across a pile of pallets at a different point it might have trouble matching the two points on its map. This is when handling dynamics becomes crucial and is a typical characteristic of modern Lidar SLAM algorithms.
Despite these challenges however, a properly designed SLAM system can be extremely effective for navigation and 3D scanning. It is particularly useful in environments that do not let the robot depend on GNSS for positioning, such as an indoor factory floor. It is crucial to keep in mind that even a well-designed SLAM system could be affected by errors. It is essential to be able to detect these flaws and understand how they impact the SLAM process to fix them.
Mapping
The mapping function creates a map of the robot's surroundings that includes the robot itself as well as its wheels and actuators as well as everything else within the area of view. This map is used for localization, path planning and obstacle detection. This is an area in which 3D lidars are particularly helpful because they can be utilized as a 3D camera (with a single scan plane).
The map building process can take some time however, the end result pays off. The ability to build an accurate, complete map of the surrounding area allows it to conduct high-precision navigation, as being able to navigate around obstacles.
As a general rule of thumb, the greater resolution the sensor, more accurate the map will be. Not all robots require high-resolution maps. For example, a floor sweeping robot may not require the same level of detail as an industrial robotics system that is navigating factories of a large size.
There are a variety of mapping algorithms that can be used with LiDAR sensors. Cartographer is a very popular algorithm that employs a two-phase pose graph optimization technique. It corrects for drift while maintaining a consistent global map. It is especially beneficial when used in conjunction with the odometry information.
GraphSLAM is a different option, that uses a set linear equations to represent the constraints in diagrams. The constraints are represented as an O matrix, and an X-vector. Each vertice in the O matrix represents a distance from an X-vector landmark. A GraphSLAM Update is a series of subtractions and additions to these matrix elements. The end result is that all O and X Vectors are updated in order to reflect the latest observations made by the robot.
SLAM+ is another useful mapping algorithm that combines odometry and mapping using an Extended Kalman filter (EKF). The EKF updates the uncertainty of the robot's position as well as the uncertainty of the features that were mapped by the sensor. The mapping function is able to make use of this information to improve its own position, which allows it to update the underlying map.
Obstacle Detection
A robot needs to be able to detect its surroundings to overcome obstacles and reach its destination. It utilizes sensors such as digital cameras, infrared scanners sonar and laser radar to determine its surroundings. In addition, it uses inertial sensors to determine its speed and position, as well as its orientation. These sensors aid in navigation in a safe manner and prevent collisions.
One important part of this process is the detection of obstacles that involves the use of a range sensor to determine the distance between the robot and the obstacles. The sensor can be mounted on the robot, inside an automobile or on the pole. It is important to remember that the sensor can be affected by a variety of elements such as wind, rain and fog. Therefore, it is essential to calibrate the sensor before each use.
The most important aspect of obstacle detection is to identify static obstacles. This can be accomplished by using the results of the eight-neighbor cell clustering algorithm. However, this method has a low accuracy in detecting due to the occlusion caused by the gap between the laser lines and the speed of the camera's angular velocity making it difficult to recognize static obstacles in one frame. To solve this issue, a method called multi-frame fusion was developed to improve the detection accuracy of static obstacles.
The method of combining roadside unit-based as well as obstacle detection using a vehicle camera has been proven to increase the efficiency of processing data and reserve redundancy for further navigational operations, like path planning. The result of this technique is a high-quality image of the surrounding area that is more reliable than a single frame. In cheapest lidar robot vacuum tests the method was compared against other methods for detecting obstacles such as YOLOv5 monocular ranging, and VIDAR.
The results of the experiment proved that the algorithm was able accurately identify the position and height of an obstacle, as well as its tilt and rotation. It also had a great ability to determine the size of obstacles and its color. The method also exhibited solid stability and reliability even when faced with moving obstacles.