Department of Automation, Chongqing University of Posts and Telecommunications, No. 2 Chongwen Road, Chongqing 40000, China.
Chongqing Changan New Energy Science and Technology Co., Ltd., Chongqing 401120, China.
Sensors (Basel). 2020 Apr 15;20(8):2238. doi: 10.3390/s20082238.
Object detection, as a fundamental task in computer vision, has been developed enormously, but is still challenging work, especially for Unmanned Aerial Vehicle (UAV) perspective due to small scale of the target. In this study, the authors develop a special detection method for small objects in UAV perspective. Based on YOLOv3, the Resblock in darknet is first optimized by concatenating two ResNet units that have the same width and height. Then, the entire darknet structure is improved by increasing convolution operation at an early layer to enrich spatial information. Both these two optimizations can enlarge the receptive filed. Furthermore, UAV-viewed dataset is collected to UAV perspective or small object detection. An optimized training method is also proposed based on collected UAV-viewed dataset. The experimental results on public dataset and our collected UAV-viewed dataset show distinct performance improvement on small object detection with keeping the same level performance on normal dataset, which means our proposed method adapts to different kinds of conditions.
目标检测作为计算机视觉的一项基本任务，已经取得了巨大的发展，但对于无人机视角来说，由于目标规模较小，仍然是一项具有挑战性的工作。在本研究中，作者针对无人机视角下的小目标开发了一种特殊的检测方法。该方法基于 YOLOv3，首先通过串联两个具有相同宽度和高度的 ResNet 单元来优化 darknet 中的 Resblock。然后，通过在早期层增加卷积操作来丰富空间信息，改进整个 darknet 结构。这两种优化都可以扩大感受野。此外，还收集了用于无人机视角或小目标检测的无人机视角数据集。还基于收集的无人机视角数据集提出了一种优化的训练方法。在公共数据集和我们收集的无人机视角数据集上的实验结果表明，在保持对正常数据集的相同性能水平的情况下，对小目标检测的性能有明显的提高，这意味着我们提出的方法适应不同的条件。
