Testing & Results

Our testing yielded promising results, as the robot successfully navigated to its designated GPS waypoints while also skillfully avoiding obstacles while maintaining a safe distance from them. Furthermore, our computer vision algorithms proved to be highly accurate in detecting landmines. To delve deeper into our subsystems' results, this section will follow a similar structure to the previous section: 

In terms of the computer vision-based landmine detection system, the AVDL was able to detect three types of landmines: VS50, PMN, and PFM-1, with an accuracy of over 90% for each landmine when using the YOLOv5 computer vision algorithm.. The system also successfully and repeatably returned an image to the operator whenever a landmine was detected.

We trained the model with image data from actual mine fields as well as customized data that was generated in similar environments with the 3D printed replicas of the landmines. The total amount of images acquired turned out to be 457 images that were all manually labeled to ensure implementation of clean data, but after augmenting the data to ensure generalization of our model, the complete size of the dataset turned out to be 752 images. The training process took 40min, which resulted in a Mean Average Precision (mAP) of 0.995 and an average loss of 0.0001931; the model converged in about 4min, which further proves the high computational speed and efficiency of the YOLOv5 model (refer to Graph 2 for graphical results of the YOLOv5 model). 

The underground landmine detection system utilized a commercial 8” coil metal detector with resulted in an accuracy of over 75%. Our testing plan involved constructing a test chamber with the ability to adjust the depth of detection, types of metals to test, and ground types. We used 100g of Steel as the metal type for our experiments and filled the chamber with fresh soil, sand, and rocks. The results showed that the system was able to detect metal in each soil type at at a maximum depth of 6in. The pictures below shows the results of underground metal detection for the soil, rocks, and sand ground types respectively, where a check mark signifies successful detection, and an ‘X’ mark signifies failed detection. And for a more analytical representation, Table 2 shows the complete results of our underground landmine detection system. 

In evaluating the terrain adaptability of our AVDL, it was found that the AVDL was able to successfully navigate over all four terrains (rocks, tall grass, roots, and sand), achieving a pass score for almost every test run (refer to Table 3 and Figure 9 for visual representation of the grass, rocks, sand, and roots terrain respectively). The maximum inclination the AVDL was able to traverse was found to be 35°, as it was also tested with other inclination angles as can be seen in Table 4. These results demonstrate the robot's ability to effectively navigate over various terrain types, enhancing its overall utility in a range of field conditions. The AVDL was able to adapt successfully to the four tested terrains.  

The field experiments with the actual robot were conducted in various outdoor environments, including open fields and parking lots. The robot was able to successfully navigate to a given GPS waypoint while avoiding obstacles and maintaining a safe distance from them. The simultaneous mapping of the GPS location of the robot and any detected landmines was also evaluated and proved successful during the outdoor testing, as can be seen in Figure 11. Deviations from the path can also be seen in the figure which represents the changes in path that the robot had to partake to avoid obstacles during the waypoint navigation.  

The results of both simulation and real-world testing demonstrated the effectiveness of the waypoint-based navigation system in enabling the robot to navigate autonomously while avoiding obstacles and detecting landmines.