Safe landing in crowded areas is a crucial task for the successful integration of Unmanned Aerial Vehicles in common civilian applications. In this work, we propose a lightweight algorithm for identifying safe landing regions to prevent hurting people in emergency landing situations. To do so, a lightweight convolutional neural network architecture is developed to perform crowd detection and counting using fewer computer resources without a significant loss on accuracy. The architecture was trained using the Bayes loss function to improve its accuracy and then pruned to further reduce the computational resources used. The proposed architecture was tested over the USF-QNRF dataset, achieving good accuracy while maintaining a competitive number of parameters, around 0.067 Million, which is suitable for real-time embedded applications. The developed lightweight model is then used to obtain the density map of the crowd in real-time, which is employed to determine the largest circular region without persons by means of the polylabel algorithm. Real-time experiments validate the proposed strategy implemented in a low-cost mobile phone application and a commercial drone.

在拥挤区域中安全着陆是将无人机成功集成到普通民用应用中的关键任务。在这项工作中，我们提出了一种轻量级算法，用于识别安全着陆区域，以防止在紧急着陆情况下伤害人员。为此，开发了一种轻量级的卷积神经网络体系结构，以使用较少的计算机资源执行人群检测和计数，而不会显着降低准确性。使用贝叶斯损失函数对体系结构进行了训练，以提高其准确性，然后对其进行修剪以进一步减少所使用的计算资源。在USF-QNRF数据集上对提出的体系结构进行了测试，不仅获得了良好的准确性，而且还保持了具有竞争力的参数数量（约0.67百万），非常适合实时嵌入式应用。然后将开发的轻量模型用于实时获取人群的密度图，该密度图用于通过多标签算法确定没有人的最大圆形区域。实时实验验证了在低成本手机应用和商用无人机中实施的拟议策略。