Abstract:
Large, unmapped puddles are common in farmenvironments. Reliably detecting and avoiding these puddles is an important problem that must be solved to spread the adoption of agricultural robots. This thesis investigates sensory techniques available that could be used to detect puddles. The focus will be on hyperspectral imaging, features of the reflected signals of sonar sensors, and proximity capacitive sensors. Hyperspectral was used for distance capabilities, while sonar and capacitive sensors were cost effective backup systems. Hyperspectral imaging used band ratio analysis to determine the material of the pixel. The wavelengths 680 nm and 750 nm showed the greatest differences between grass and water. The band ratio ofwater and grasswere approximately 1.43, and 0.11. A monochrome camera with filters at 680 nm and 750 nm was used as an alternative option. Similar results were obtained, where the band ratio of water and grass were 0.98 and 0.84. Sonar sensors used its differing voltage response to distinguish different ground material. It was observed that grass was seen as a soft object that deflected the sonar pulse, whereas water had a stronger reflection back. At a height of 7 cm the water had an area value of 2.4 Vms, while grass had 1.2 Vms. Machine learning was implemented to automate the classification of sonar data with a 72.7% accuracy. The effect of rotating the sonar sensor was studied to determine the maximum range. Results showed that the materials were hard to distinguish. Simulations were used to model the proximity capacitive sensors. The results highlight that the capacitance for water is higher than soil due to its higher dielectric constant. Because of this difference, a prototype was built. However, the results from this showed little to no trend, meaning that background capacitance was detected as opposed to water or soil.