2.2. Data Acquisition and Processing

The LiDAR was connected to an ADLINK embedded computer through an Ethernet cable. The Sick Open Portal for Applications and Systems (SOPAS) Engineering Tool software was used to control the sensor and target detectability when the LiDAR was operating. The angle of the LiDAR (below horizontal) was altered before each scan, to ensure the maximum visibility of the rods in SOPAS.

The LiDAR can be controlled via the Robotic Operation System (ROS) open-source robot management software [32]. The produced ROSBAG files from the LiDAR were stored on the computer. A bag file stores ROS message data. The recorded ROSBAG files were processed in the ROS environment [35]. The GPS location of the LiDAR and the recorded point clouds during the scans were extracted from the ROSBAG files. The point clouds included the xyz coordinates and the reflected intensity values of each point. Point clouds were visualized in CloudCompare software version 2.12 alpha and the point cloud for the rods/weeds was segmented and separated by creating a moving window and each rod/weed studied individually. The moving window was aligned with the x, y, z axes and moved in all three dimensions during the segmentation. By creating a moving window, the x, y, z coordinates of rods were extracted from the point cloud. The z values were compared to the rods’ actual heights to confirm the objects were selected accurately. The actual height of the rods/weeds were compared with the estimated LiDAR data for the ‘suspected’ rods/weed to assess the accuracy of the estimated height of the targets.

Clustering was performed on the normalized reflected values from targets to understand the capability of discriminating the targets based on their return values. The reflected values in a small xy region where two different targets existed (lawn vs. rod or wheat vs. weeds) were clustered to see if two distinct clusters were detected. Clustering is one of the unsupervised data mining methods, which deals with grouping the data based on the similarity of the objects in each group cluster [36]. Silhouette measurements can be used to understand the performance of the clustering algorithms. The correctness of the assignment of the data to a particular cluster can be measured by the silhouette score. Silhouette score values vary between −1 to +1, where +1 value shows the object has been clustered correctly and −1 shows the object has not been properly clustered [37].

The GPS coordinates of each target rod were recorded before each scan. The estimated GPS coordinates of each rod in trial 1 was calculated based on the GPS coordinates of the LiDAR at each scan and the estimated x and y coordinates of each rod. The accuracy of the estimated location of each rod was assessed by calculating the Euclidean distance [38]. The Euclidean distance was measured between the estimated position and the actual GPS coordinates of each rod using the Sklearn package in Python 3.7.