基于多传感器融合的无人机生命搜寻方法

UAV Life Search Method Based on Multi-sensor Fusion

为应对单个生命探测传感器在野外、灾区生命搜寻时的不稳定状况,提出了一种基于多传感器信息融合的无人机(UVA)生命搜寻方法。首先,构建不同结构的ResNeXt网络以提取不同维度信息的特征,利用一维ResNeXt网络提取音频梅尔频谱系数的深层特征,利用二维ResNeXt网络提取红外图像的深层特征;其次,使用判别相关分析(DCA)对2种高维特征进行降维融合,兼顾不同特征的相关性和类别性,以获得更丰富的环境信息,从而提高生命搜寻的准确性;最后,将融合特征输入支持向量机分类器以进行生命识别的决策,建立具有相关性的音频和图像双模态数据集,并将所提方法在该数据集上进行实验比较和分析,对其性能进行有效评估。实验结果表明:所提方法在特征提取和特征融合方面效果优于其他传统方法,且多传感器融合识别准确率可达98.7%,证明该方法能有效提高特殊场景下人体检测的准确性,多传感器融合检测效果优于单传感器。

In order to cope with the unstable conditions of single sensor for life detection in the fields and disaster areas, a life search method was proposed based on multi-sensor information fusion. Firstly, ResNeXt networks with different structures were constructed to extract features with different dimensional information. Deep features of audio Mel-scale Frequency Cepstral Coefficients were extracted using a one-dimensional ResNeXt network, and deep features of the infrared images were extracted using a two-dimensional ResNeXt network. Secondly, the two high-dimensional features were fused by dimensionality reduction using discriminant correlation analysis(DCA) to take into account the correlation and category of different features in order to obtain richer environmental information, thus improving the life search accuracy. Finally, the fused features were fed into a support vector machine classifier for decision making in life recognition. A bimodal dataset of audio and images with correlation was created and the proposed method was experimentally compared and analyzed in this dataset to evaluate the search performance of the proposed method. The experimental results demonstrated that the proposed method could outperform other traditional methods in feature extraction and feature fusion, and the multi-sensor fusion recognition accuracy could reach 98.7%, which proved that the method could effectively improve the accuracy of human detection in special scenes, and the performance of multi-sensor fusion based human detection was higher than that with single sensor.