利用无人机航拍视频结合YOLOv3模型和SORT算法统计云杉数量

准确、快速地统计苗木数量对苗圃的运营和管理具有重要意义，是提高苗圃运营和管理水平的有效方式。为快速准确统计完整地块内苗木数量，该研究选取云杉为研究对象，以无人机航拍完整地块云杉视频为数据源，提出一种基于YOLOv3（You Only Look Once v3，YOLOv3）和SORT（Simple Online and Realtime Tracking，SORT）的云杉数量统计方法。主要内容包括数据采集、YOLOv3检测模型构建、SORT跟踪算法和越线计数算法设计。以平均计数准确率（Mean Counting Accuracy，MCA）、平均绝对误差（Mean Absolute Error，MAE）、均方根误差（Root Mean Square Error，RMSE）和帧率（Frame Rate，FR）为评价指标，该方法对测试集中对应6个不同试验地块的视频内云杉进行数量统计的平均计数准确率MCA为92.30%，平均绝对误差MAE为72，均方根误差RMSE为98.85，帧率FR 11.5 帧/s。试验结果表明该方法能够快速准确统计完整地块的云杉数量。相比SSD+SORT算法，该方法在4项评价指标优势显著，平均计数准确率MCA高12.36%，帧率FR高7.8 帧/s，平均绝对误差MAE和均方根误差RMSE分别降低125.83和173.78。对比Faster R-CNN+SORT算法，该方法在保证准确率的基础上更加快速，平均计数准确率MCA仅降低1.33%，但帧率FR提高了10.1 帧/s。该研究从无人机航拍视频的角度为解决完整地块的苗木数量统计问题做出了有效探索。

Quantity statistics of spruce under UAV aerial videos using YOLOv3 and SORT

Abstract: A seedling quantity is a key indicator to predict the actual production, supply, and demand for the operation and management of a nursery. The manual visualization has still dominated the statistics for the number of seedlings in complete plots. However, the application needs cannot be fully met in recent years, such as high cost, low efficiency, and slow data update. Therefore, it is necessary to fast and accurately estimate the number of seedlings in the whole plots. Taking the spruce as the research object, this study aims to propose a quantity statistics approach under Unmanned Aerial Vehicle (UAV) aerial videos using YOLOv3 and SORT. The specific procedure included the data acquisition, YOLOv3 detection model, SORT tracking, and cross-line counting. Two areas were divided for the image and video acquisition, each with 6 complete test plots. In the stage of data acquisition, 558 images and 6 videos were captured by a DJI Phantom 4 (UAV). The quantity statistics dataset was then constructed with the acquired images and videos, where the training dataset contained 558 images, and the test dataset contained 6 videos. Subsequently, a YOLOv3 model was selected to detect the spruce, while a SORT model was to track the spruce, and the cross-line counting to count the number of spruce. The performance of the combined YOLOv3+SORT was also quantitatively evaluated using Mean Count Accuracy (MCA), Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and Frame Rate (FR). It was found that the MCA of 92.30%, MAE of 72, RMSE of 98.85, and FR of 11.5 frames/s for the test dataset in the quantity statistics. The experimental results showed that quick and accurate counting was achieved for the number of spruce in the complete plots. The YOLOv3+SORT was also compared with the SSD+SORT and Faster R-CNN+SORT, in order to further verify the performance of the model. The results show that the YOLOv3+SORT performed over the SSD+SORT in all four evaluation indexes. Particularly, the YOLOv3+SORT was much faster with higher guaranteed accuracy, with 1.33% lower MCA, and 10.1 frames/s higher FR, compared with the Faster R-CNN+SORT. In summary, the quantity statistics using YOLOv3 and SORT can be widely expected to serve as an effective way to rapidly and accurately count the number of seedlings in the whole plots. This study can also offer promising potential support to the seedling quantity statistics from the perspective of UAV aerial videos.