An Evaluation of Three Different Infield Navigation Algorithms An Evaluation of Three Different Infield Navigation Algorithms

In this chapter, we present and evaluate three different infield navigation algorithms, based on the readings from a LIDAR sensor. All three algorithms are tested on a small field robot and used to autonomously drive the robot between the two adjacent rows of maze plants. The first algorithm is the simplest one and just takes distance read ings from the left and right side. If robot is not in the center of the mid-row space, it adjusts its course by turning the robot in the right direction accordingly. The second approach groups the left and right readings into two vertical lines by using least-square fit approach. According to the calculated distance and orientation to both lines, it adjusts the course of the robot. The third approach tries to fit an optimal triangle between the robot and the plants, revealing the most optimal one. Based on its shape, the course of the robot is adjusted. All three algorithms are tested in a simulated (ROS stage) and then in an outdoor (maze test field) environment comparing the optimal line with the actual calculated position of the robot. The tests prove that all three approaches work with an error of 0.041 ± 0.034 m for the first algorithm, 0.07 ± 0.059 m for the second, and 0.078 ± 0.055 m error for the third.


Introduction
The usual approach of autonomously driving agriculture machinery trough the field is by using a precise differential [1] or RTK global positioning systems [2]. In order to work, these systems require a previously known path of movement that is repeated with each iteration when the crop needs to be treated [3]. This is not always possible, if no prior GPS information exists, or simply because an accurate GPS system is not always available. So, different systems should be used in these situations.
A possible solution to solve this is to use cameras to detect plant lines [4] or even stereo cameras to build a 3D cloud of points [5], describing the plants, and drive the agricultural machinery to drive between them. Another possible approach is the use of LIDAR systems [6] that take accurate measurements of the scene, by reading distances from the sensor to the first obstacle and repeating this for the whole range (usually 270° or 360° in 1°, 0.5° or even 2.25° steps).
With the help of these systems, the machines [7][8][9] can drive even in an unfamiliar field where they have not been applied before. They rely on the property of the field and the crops planted. The plants are in parallel to each other, and the machines can drive between the crop lines, in mid-row spaces, in order to not damage the plants. This can be done by using different approaches/algorithms to guide the machines.
More advance systems using different SLAM methods [10,11] even build a map of the environment and localize the machines in a new environment that is only being discovered. Based on the map, it then constructs a path using path planner [12] and path follower [13] to follow the path. These approaches come in use in case of unfamiliar scenes but are not really necessary in infield situations where prior information regarding the pattern of the plants is known.
So, the purpose of this work is to first present three different algorithms that could be used to autonomously drive the machines and to evaluate their accuracy for infield navigation. All three were implemented as part of robotic operating system (ROS) [14] and applied on a small field robot, making it autonomous when driving through the field in mid-row spaces. Their purpose is not to plan and follow the path but to adjust the heading of the machines/robot at every measured location. The field robot that was used to test the algorithms is presented in Section 2, along with all three algorithms used, which are then evaluated in Section 3.

Materials and methods
The algorithms are tested on a small field robot [16] depicted in Figure 1. It is an electric driven robot with 50 × 50 cm in size, small enough to fit in 75 cm wide space between two parallel lines of maze plants. It has four in-wheel BLDC motors capable of delivering 200 W max peak power, equipped with four additional motors to individually turn each of the wheels, making possible to drive in different steering modes like skid steer or Ackermann. Besides the odometry from the wheels, the robot is equipped with three different additional sensors: two digital cameras, an IMU unit, and a LIDAR sensor. The data from the sensors are processed by two onboard computers, a low-level computer build around Raspberry PI 3B and Intel NUC i7 (gen. 7) computer, used as high-level processing unit. The two units run a Linux-based distribution with robotic (meta) operating system (ROS) on top, configured in multimachine mode to split the essential processing from time consuming, advance algorithms, making the robot as responsive as possible.
If the robot, as the one in Figure 1, has to drive autonomously through the crop fields, it must have a navigation algorithm. The navigation algorithm relies on the distance measurements from the LIDAR sensor. The experimental robot is equipped with the SICK TIM310 LIDAR, with a 270° area at a 1° angular resolution. It detects obstacles up to 4 m away and then chooses the optimal path based on the readings. Once the robot reaches the end of the crop line, it uses data from the on-board compass (IMU unit) and turns in the next row.
In this chapter, we investigate the accuracy of three different navigation algorithms and compare them, with left and right row distance as a reference for optimal path that should be in the middle of the rows and with as little oscillations as possible.

Minimal row offset-based algorithm
The first algorithm is the simplest one. It takes 30 readings from left and 30 from right side, as shown in Figure 2.
From these two sets, it first eliminates ones that are too far away, that is more than 0.75 cm and belong to the other crop lines, and then calculates an average distance value for each side. These average distances can be written as d r and d l for right and left side, respectively. These two are then used to calculate an offset, as shown in Eq. 1: The value of the offset is then used to adjust the course of the robot, if necessary, as shown in Eq. 2.
So, the current course of the robot is adjusted by the new value orientation which also corresponds directly how much the wheels should turn in radians.

Least-square fit approach algorithm
This second algorithm was designed to navigate the robot between two walls that are parallel to each other, this being either an artificial barrier, e.g., walls, or real crop lines, such as maze plants. The overview of the approach is depicted in Figure 3, explaining each sensingadjustment cycle.
(A) The sensor reads the data-the distances between the sensor and the obstacles for each degree-and triggers a callback function each time it completes the measuring sequence.
(B) The callback function first filters the data depending on the distance readings. The points that are too far away and points that are too near to count are discarded. The algorithm makes possible to set how many points should be included for each count for each side that corresponds to how many degrees will use in the subsequent steps of the algorithm. The useful readings are stored in two data sets, one for left and one for right side.  (C) In the third step, a linear fit is used, for which a least squares method [17] was chosen. This way the slope and y-intercept of a linear equation describing each set for each side is calculated.
The least squares method allows us to linearly fit the measurements with a smaller number of heavy duty mathematical operations. For this, we need to define some additional parameters with which we then calculate the slope and y-intercept of each line: x sum − sum of all the distances taken for the line , and distances x L and y L can be calculated using Eqs. (5) and (6): (E) Once the information about the intersection from the two lines is known, the position of the robot is calculated. The distance y L stays constant, if the robot is aligned up with the row no matter which wall it is closer. The x L describes how far away it is from both of walls/lines. With just looking at the two distances, the problem is simplified and can be solved with Eqs. (1) and (2).
Based on the position of the robot, the described approach can describe two different situations. The first situation is when the robot is in the right location, as shown in Figure 4, and the second if the robot is not in the right location and needs adjustment, as shown in Figure 5. In both cases, the parameters of the two linear equations and the distance to the path of the robot are used.
Since the walls are always apart from each other with a constant width, the triangle covers the same surface. What changes is the orientation and position changes with the robot that produces different triangles. If the robot is not aligned in parallel to the walls or crop lines, an asymmetric triangle is constructed, as the one in Figure 5.

Triangle-based navigation algorithm
The third approach [15] of finding an optimal path for the robot consists out of multiple steps. As shown in Figure 6, the algorithm uses trigonometric functions to calculate the distance of a segment between every two sequential points from LIDAR sensor. The segment distance must be wide enough to drive the robot trough, and if they do not meet the criteria, they are disposed. This is depicted in Figure 6 where red colored segment distances written as (d 1 , d 5,     Agricultural Robots -Fundamentals and Applications Figure 6 shows that only segment d 7 is appropriate for robot to drive trough. The procedure eliminates inappropriate segments as shown in Figure 7. Algorithm calculates the angles: α, β, and γ in triangle limited between two sequential points and LIDAR sensor. If any of the angles α or β is bigger than predefined threshold, set to 100° or more, which would produce a triangle that would not fit in field situations, the segment is disposed. Disposed segments are marked with red triangle in Figure 7, and the optimal segment d 7 which pass the criteria is marked as green triangle.
When the optimal segment is determined, the algorithm calculates the angle for the front wheels to turn. The midpoint of optimal segment is calculated for robot to drive to. Angle δ in Figure 8 represents the angle for robot to turn front wheels to follow the midpoint of optimal segment.

Results
In order to test and compare all three algorithms from Section 2, two separate approaches are used. In the first, a simulated environment is built to test all three using precisely the same data sets in order to verify if they work. In the second approach, a real testing environment is used in order to evaluate the approaches in uncontrolled environment.
In contrast to the first experiment, where the algorithms were tested one time, the robot in the second approach is tested using one of three algorithms and repeated five times, for example, it drives between the two rows five times, in order to calculate an average absolute value with its standard deviation. The average value is computed for all three algorithms and evaluated as the average displacement from the center point and its standard deviation. The algorithm that preforms best should have the average closest to the real middle point, with as little deviation as possible, corresponding to a minimal oscillation pattern. Figure 9 depicts an exaugurated principle of how the robot moves. The oscillation pattern is of course small and neglectable with some algorithms.

Simulated environment
A simulated environment was build using ROS stage simulator [18] in order to test all three algorithms before the application in real environment. In this experiment, all tests start at the same starting point, and the robot moves in semi-parallel direction to two artificial walls, mimicking two plant rows (Figure 10).

Real environment
In the second experiment, the robot moves between two plant rows. In this experiment, the data sets are not precisely the same as the plants can move due to wind. Even in no wind situation, the data sets might differ due to quantization steps of the LIDAR that might not be at the same location each time. Figure 11 depicts the environment used in this approach. The results of the second test are presented in Table 1, where in each iteration, the average distance from the mid-row path was calculated on 8 m long test runs, but this time using real plants in real environment.

Agricultural Robots -Fundamentals and Applications
The results in Table 1 show that the best performing algorithm is the algorithm from Section 2.1 with an error of 0.041 ± 0.034 m, and the second and third are very close with an average error of 0.07 ± 0.059 m for the second from Section 2.2 and 0.078 ± 0.055 m error for the third from Section 2.3. It should be noted that all three performed well reaching the end of each row without any problems. The difference between the first and other two is that the first uses the values in small proximity to the sensor on either side, but the second and third use a bigger range on both sides making them more useful in situations when rows are not straight as in

Conclusion
In this chapter, we presented three different algorithms for infield navigation and then tested them using first a simulated environment and, second, evaluated them in a real environment. The results from Section 3 show that all three algorithms perform good, with the best one in terms of optimal mid-row driving and minimal oscillation, the algorithm from Section 2.1 with an error of 0.041 ± 0.034 m, the second being algorithm from Section 2.2 with an error of 0.07 ± 0.059 m, and third, the algorithm from Section 2.3 with an error of 0.078 ± 0.055 m.
The algorithms presented in this chapter adjust the movement of the robot according to each iteration of measured distance sets recorded by LIDAR sensor. As a logical improvement to the best performing algorithms from Section 2, the accuracy could be improved by taking into account the measurements further away from the robot, measurements of the parallel lines, that are currently filtered out, as well as measurements of the previously driven row(s), as the rows are always in parallel to each other in which we get a prior information for the current rows. All this would further improve the performance of the algorithms. Agricultural Robots -Fundamentals and Applications