Developing a High Precision Cotton Boll Counting SystemUsing Active Sensing

摘要

Different sensing technologies, such as laser scanning or stereoscopic reconstruction, were introduced for 3D imaging of plants in field conditions. Because of the limitations of imaging technologies, capturing images from an appropriate distance canprovide better resolutions. Traditionally, most studies have used fixed sensors on unmanned mobile platforms, and then automatically guided the scan through the field to capture images. Since the range of unmanned mobile platforms in the field is often limited by plants, it is difficult to maintain the quality of captured images. Thus, an active sensing technique is necessary to ensure the quality of the captured images. In this study, a Kinect v2 sensor was mounted on a robotic arm with 6DOF to captureimages around cotton plants. Both color and depth images were captured by Kinect. A deep learning model was used to detect cotton bolls from color images with a bounding box. Next, the detected bounding boxes were used to segment point clouds and estimate the position of the detected cotton boll. All estimated positions were then transferred to world coordinates to remove duplicates with a threshold of distance. Last, the number of cotton bolls was counted, and it was determined that the accuracy of this system can reach 95.56%.