Effect of Camera Shutter Mechanism on the Accuracy of a Custom YOLOv8 Model for Pattern Recognition in Motion on a UGV
James Kemeshi, Mohammad Ashik Alahe, Young Ki Chang, Pappu Kumar Yadav · 2024
Abstract. Facilitating the adoption of unmanned ground vehicles (UGVs) by small-scale farmers (SSFs) is an important step to achieving sustainable agriculture in the USA. To do this, we commenced the Reduction-To-Below-Two (R2B2) grand project that envisions addressing the high cost of commercially available agricultural UGVs through cost reduction. The R2B2 project offers a solution by developing a UGV with a production cost below $2000 USD. During this project, we assessed the performance of a custom YOLOv8 model in a pattern recognition task using two cameras, a global shutter camera (AR0234) and a rolling shutter camera (IMX462) that cost USD 110 and USD 40, respectively. The specific purpose of this experiment was to determine the feasibility of using a low-cost camera for data acquisition and crop health monitoring. In this work, we assessed the effect of varying environmental conditions on the model's performance with both cameras in an indoor experiment that follows a 4 x 3 x 2 factorial experiment with three replications. We compared the effect of varying speeds of the R2B2-UGV on which both cameras were mounted (0.75 m/s, 1.0 m/s, 1.25 m/s, and 1.50 m/s), illumination incident on the surface of the target (100 Lux, 650 Lux, and 1250 Lux), and terrain conditions (smooth and undulating) on the accuracy of the model to detect four classes of patterns. To reduce the effect of the fisheye lens of the IMX462 camera, we filtered the dataset to include only frames with centrally aligned targets. Results revealed that the model had an average precision percentage of 85.39% and 80.07% with the IMX462 and AR0234, respectively. In contrast, the results showed that the model had an average recall percentage of 92.37% and 86.93% with the AR0234 and IMX462, respectively. Overall, the model had F1 scores of 85.35% and 85.53% with the AR0234 and IMX462, respectively. This result affirms our hypothesis that we could use the IMX462 camera for machine vision applications in similar environmental conditions. This result does not represent the model's performance on both cameras over the entire dataset. However, it shows that there is a possibility of using rolling shutter cameras in certain conditions. A recommendation for future work would be to include all frames from each treatment with the target in view in the dataset to evaluate the model's performance with both cameras to provide a more comprehensive analysis.