Click on “Download PDF” for the PDF version or on the title for the HTML version.

If you are not an ASABE member or if your employer has not arranged for access to the full-text, Click here for options.

Development and Test of an RGB-D Camera-based Rock Detection System and Path Optimization Algorithm in an Indoor Environment

Published by the American Society of Agricultural and Biological Engineers, St. Joseph, Michigan

Citation:  2021 ASABE Annual International Virtual Meeting  2100105.(doi:10.13031/aim.202100105)
Authors:   Jithin Jose Mathew, Yu Zhang, Paulo Flores, Cannayen Igathinathane, Zhao Zhang
Keywords:   rock picker, rock detection, RGB-D, data fusion, path optimization

Abstract. The presence of rocks on the soil surface of agricultural fields can be a severe problem for farmers. This scenario worsens in the northern parts of the U.S with prolonged winter, where freezing and thawing of water in the soil tend to push rocks upward to surface. In addition to affecting crops (e.g., germination rate), large and dense rocks can damage agricultural machinery, such as tillage equipment blades, combine headers, and hay balers, leading to costly fixes and potentially delaying field operations. Farmers from the state of North Dakota, especially those growing crops that require to operate combine headers close to the ground surface, tend to be affected more seriously by this issue. The current labor-dependent rock removal method is inefficient and has the potential to cause occupational injuries. Therefore, the need for an automatic system that can pick rocks is on the rise. This study developed and tested an RGB-D (red, green, blue, and depth) camera-based innovative approach to detect rocks. The RGB-D image and depth data are integrated to detect and locate rocks in real-time mode. After rocks are detected, path optimization was conducted to pick multiple rocks with minimal time (short total traveling distance). Several path configuration algorithms were proposed and compared to determine the satisfactory path. The system was designed and integrated in indoor experimental. This study evaluates the YOLOv3-tiny object detection model before and after retraining on augmented data. Our evaluation showed that retrained YOLOv3-tiny model shows improved performance resulting in a mAP@.5 of 0.91 and F1-Score of 0.85. The methods developed in this study is a significant contribution to precision agriculture and can be scaled and applied for rock picking. Ongoing and future work should focus of improving model performance and compare different DL architectures.

(Download PDF)    (Export to EndNotes)