LiDAR Robot Navigation
LiDAR robot navigation is a complex combination of mapping, localization and path planning. This article will present these concepts and demonstrate how they work together using an example of a robot achieving its goal in a row of crops.
LiDAR sensors are relatively low power requirements, allowing them to increase the battery life of a robot and decrease the amount of raw data required for localization algorithms. This allows for a greater number of variations of the SLAM algorithm without overheating the GPU.
LiDAR Sensors
The heart of lidar systems is their sensor, which emits pulsed laser light into the environment. The light waves hit objects around and bounce back to the sensor at various angles, depending on the structure of the object. The sensor is able to measure the time it takes to return each time, which is then used to calculate distances. The sensor is usually placed on a rotating platform, which allows it to scan the entire area at high speeds (up to 10000 samples per second).
LiDAR sensors are classified based on their intended airborne or terrestrial application. Airborne lidars are typically attached to helicopters or UAVs, which are unmanned. (UAV). Terrestrial LiDAR systems are generally mounted on a stationary robot platform.
To accurately measure distances, the sensor needs to be aware of the exact location of the robot at all times. This information is recorded using a combination of inertial measurement unit (IMU), GPS and time-keeping electronic. These sensors are used by LiDAR systems in order to determine the exact location of the sensor in space and time. This information is used to build a 3D model of the environment.
LiDAR scanners can also detect various types of surfaces which is especially useful when mapping environments that have dense vegetation. For instance, if a pulse passes through a forest canopy, it will typically register several returns. The first one is typically attributed to the tops of the trees, while the last is attributed with the ground's surface. If the sensor captures each peak of these pulses as distinct, this is referred to as discrete return LiDAR.
Discrete return scanning can also be helpful in analysing the structure of surfaces. For instance, a forested region might yield a sequence of 1st, 2nd, and 3rd returns, with a final large pulse representing the bare ground. The ability to separate and store these returns as a point cloud allows for detailed models of terrain.
Once a 3D map of the surroundings has been created and the robot is able to navigate using this information. This involves localization as well as creating a path to reach a navigation "goal." It also involves dynamic obstacle detection. This is the process that detects new obstacles that were not present in the map that was created and then updates the plan of travel in line with the new obstacles.
SLAM Algorithms
SLAM (simultaneous mapping and localization) is an algorithm which allows your robot to map its surroundings, and then determine its location in relation to the map. Engineers utilize this information for a variety of tasks, including planning routes and obstacle detection.
For robot with lidar to work the robot needs a sensor (e.g. a camera or laser) and a computer that has the appropriate software to process the data. Also, you need an inertial measurement unit (IMU) to provide basic positional information. The result is a system that can accurately track the location of your robot in an unspecified environment.
The SLAM process is complex and many back-end solutions are available. No matter which solution you select for a successful SLAM is that it requires constant communication between the range measurement device and the software that extracts the data, as well as the robot or vehicle. This is a highly dynamic procedure that can have an almost unlimited amount of variation.
As the robot moves, it adds scans to its map. The SLAM algorithm then compares these scans to the previous ones using a method called scan matching. This aids in establishing loop closures. The SLAM algorithm is updated with its robot's estimated trajectory when a loop closure has been discovered.
Another factor that complicates SLAM is the fact that the surrounding changes as time passes. For instance, if your robot is walking down an empty aisle at one point and is then confronted by pallets at the next point it will be unable to finding these two points on its map. Handling dynamics are important in this case, and they are a part of a lot of modern Lidar SLAM algorithms.
SLAM systems are extremely efficient in 3D scanning and navigation despite the challenges. It is especially useful in environments that do not allow the robot to rely on GNSS-based positioning, such as an indoor factory floor. However, it is important to remember that even a properly configured SLAM system may have errors. It is essential to be able recognize these issues and comprehend how they impact the SLAM process in order to correct them.
Mapping
The mapping function builds a map of the robot's surroundings, which includes the robot as well as its wheels and actuators as well as everything else within its view. This map is used to aid in the localization of the robot, route planning and obstacle detection. This is an area where 3D lidars can be extremely useful since they can be used like the equivalent of a 3D camera (with one scan plane).
The map building process may take a while however the results pay off. The ability to build a complete and consistent map of the environment around a robot allows it to navigate with great precision, and also around obstacles.
As a rule of thumb, the higher resolution the sensor, more precise the map will be. However there are exceptions to the requirement for high-resolution maps: for example floor sweepers might not need the same degree of detail as a industrial robot that navigates factories of immense size.
There are many different mapping algorithms that can be used with LiDAR sensors. Cartographer is a well-known algorithm that employs the two-phase pose graph optimization technique. It corrects for drift while maintaining an unchanging global map. It is particularly useful when paired with the odometry information.
Another alternative is GraphSLAM which employs a system of linear equations to represent the constraints in graph. The constraints are represented as an O matrix, as well as an vector X. Each vertice of the O matrix contains the distance to a landmark on X-vector. A GraphSLAM update consists of a series of additions and subtraction operations on these matrix elements, with the end result being that all of the X and O vectors are updated to account for new robot observations.
SLAM+ is another useful mapping algorithm that combines odometry with mapping using an Extended Kalman filter (EKF). The EKF changes the uncertainty of the robot's location as well as the uncertainty of the features recorded by the sensor. The mapping function will make use of this information to improve its own position, which allows it to update the base map.
Obstacle Detection
A robot needs to be able to sense its surroundings so it can avoid obstacles and get to its desired point. It makes use of sensors such as digital cameras, infrared scanners, laser radar and sonar to sense its surroundings. Additionally, it utilizes inertial sensors to measure its speed, position and orientation. These sensors help it navigate without danger and avoid collisions.
A key element of this process is obstacle detection that consists of the use of sensors to measure the distance between the robot and obstacles. The sensor can be positioned on the robot, inside an automobile or on the pole. It is crucial to keep in mind that the sensor can be affected by a variety of elements, including wind, rain, and fog. It is important to calibrate the sensors prior to 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 is not very effective in detecting obstacles due to the occlusion caused by the distance between the different laser lines and the angle of the camera which makes it difficult to recognize static obstacles within a single frame. To overcome this problem, a technique of 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 by a vehicle camera has been proven to increase the efficiency of data processing and reserve redundancy for future navigational tasks, like path planning. The result of this technique is a high-quality picture of the surrounding environment that is more reliable than a single frame. In outdoor comparison experiments the method was compared to other methods of obstacle detection like YOLOv5 monocular ranging, and VIDAR.
The results of the test showed that the algorithm was able accurately identify the position and height of an obstacle, in addition to its tilt and rotation. It was also able to identify the color and size of an object. The method was also robust and stable even when obstacles moved.