| 摘 要： |
Human recognition in a variety of environments is a hot area of research. The results of such research are already assisting in the analysis and prediction of human behaviour based on motion, colour map video sequences, infrared video, etc. However, the physical presence of humans in such operations is often dangerous due to the inaccessibility of search areas. To solve this problem, a multi-level drone system has been proposed, where each level has its own tasks and capabilities. The system includes four levels: swarm cloud, swarm operator, swarm mate and swarm node. A human recognition model has been proposed as a complex model consisting of a convolutional neural network (CNN) and a deep learning network (LSTM). The sequential use of these networks allows both spatial (CNN) and temporal (LSTM) features to be considered. As a result, a convolutional model with long short-term memory is proposed. The solution to the problem of increased computational complexity caused by the use of multiple neural networks lies in the architectural features of a multi-level drone system. Namely, swarm nodes can exchange heterogeneous information in real time. Such communication makes it possible to narrow down the searched area to that in which the potentially searched object has already been detected. This approach implies a reduction in the dimensionality of the video sequence at the processing stage, which in turn reduces both the resources required and the time spent on recognition. The application of this model is appropriate in systems where recognition time can be sacrificed for accuracy. For instance, in a drone swarm system for search and rescue of disaster victims. The swarm can always return to the area where cloud-level calculations indicate a person is located.
在多种环境中的人类识别是当前研究的热点。相关研究成果已可用于基于运动、彩色地图视频序列、红外视频等对人类行为进行分析和预测。然而，由于搜索区域难以到达，人类在这些操作中的实际存在往往存在安全隐患。为解决这一问题，提出了一种多层级无人机系统，各层级具有独立的任务与功能。该系统包含四个层级：蜂群云、蜂群操作员、蜂群伙伴和蜂群节点。同时，提出了一种复杂的人类识别模型，由卷积神经网络（CNN）和深度学习网络（LSTM）组成。通过将这两种网络按顺序结合使用，可同时考虑空间特征（CNN）和时间特征（LSTM）。因此，提出了一种结合长短期记忆的卷积模型。多层级无人机系统的架构特性解决了多个神经网络引入的计算复杂性增加问题，具体体现在蜂群节点能够实时交换异构信息。这种通信方式能够将搜索范围缩小到已检测到潜在目标的区域。该方法在处理阶段降低了视频序列的维度，从而减少了所需资源和识别所花费的时间。这种方法适用于可以牺牲识别时间以换取准确性的系统中，例如用于搜救灾难受害者的无人机群系统。无人机群可始终返回到云层计算显示有人所在的位置区域。 |
京公网安备 11010202009788