Deep Learning based Real-time GPU-accelerated Tracking and Counting of Cotton Bolls under Field Conditions using a Moving Camera

Abstract

<b><sc>Abstract.</sc></b> <b>Robotic harvesting involves navigation and environmental perception as first operations before harvesting of the bolls can commence. Navigation is the distance required for a harvester‘s arm to reach the cotton boll while perception is the position of the boll relative to surrounding environment. These two operations give a 3D position of the cotton boll for picking and can only be achieved by detection and tracking of the cotton bolls in real-time. It means detection, tracking and counting of cotton bolls using a moving camera allows the robotic machine to harvest easily. GPU-accelerated deep neural networks were used to train the convolution networks for detection of cotton bolls. It was achieved by using pretrained tiny yolo weights and DarkFlow, a framework which translates YOLOv2 darknet neural networks to TensorFlow. A method to connect tracklets using vectors that are predicted using Lucas-Kanade algorithm and optimized using robust L-estimators and homography transformation is proposed. The system was tested in defoliated cotton plants during the spring of 2018. Using three video treatments, the counting performance accuracy was around 93% with standard deviation 6%. The system average processing speed was 21 fps in desktop computer and 3.9 fps in embedded system. Detection of the system achieved an accuracy and sensitivity of 93% while precision was 99.9% and F1 score was 1. The Tukey‘s test showed that the system accuracy and sensitivity was the same when the plants were rearranged. This performance is crucial for real-time robot decisions that also measure yield while harvesting.</b>