一种基于DSmT和HMM的序列飞机目标识别算法  

A Target Recognition Algorithm for Sequential Aircraft Based on DSmT and HMM

作　　者：李新德[1] 潘锦东[1] DEZERT Jean 

机构地区：[1]东南大学自动化学院复杂工程测量与控制教育部重点实验室,南京210096 [2]法国国家航天航空研究中心帕莱索F-91761法国

出　　处：《自动化学报》2014年第12期2862-2876,共15页Acta Automatica Sinica

基　　金：国家自然科学基金(60804063,61175091);江苏省“青蓝工程”资助计划,航空基金(20140169002);江苏省“六大高峰人才”资助计划资助~~

摘　　要：针对姿态多变化的飞机自动目标识别中的低识别率问题,提出了一种基于DSm T(Dezert-Smarandache theory)与隐马尔可夫模型(Hidden Markov model,HMM)的飞机多特征序列信息融合识别算法(Multiple features and sequential information fusion,MFSIF).其创新性在于将单幅图像的多特征信息融合识别和序列图像信息融合识别进行有机结合.首先,对图像进行二值化预处理,并提取目标的Hu矩和轮廓局部奇异值特征;然后,利用概率神经网络(Probabilistic neural networks,PNN)构造基本信度赋值(Basic belief assignment,BBA);接着,利用DSm T对该图像的不同特征进行融合,从而获得HMM的观察值序列;再接着,利用隐马尔可夫模型对飞机序列信息融合,计算观察值序列与各隐马尔可夫模型之间的相似度,从而实现姿态多变化的飞机目标自动识别;最后,通过仿真实验,验证了该算法在飞机姿态发生较大变化时,依然可以获得较高的正确识别率,同时在实时性方面也可以满足飞机目标识别的要求.另外,在飞机序列发生连续遮挡帧数τ≤6的情况下,也具有较高的飞机目标正确识别率.For the low rate problem of recognition of automatic target recognition of aircraft caused by the great change of posture, a target recognition algorithm based on DSmT （Dezert-Smarandache theory） and HMM （Hidden Markov model） is proposed by utilizing the fusion of multiple features and sequential information （i.e. MFSIF）. The novelty of the algorithm is integrating the multiple feature fusion recognition with the sequential images fusion recognition. Firstly, the sequential images are preprocessed with the binarization method, then the features of Hu moments and partial singular values of outline are picked up. Secondly, the PNN （Probabilistic neural network） is used to construct the BBA （Basic belief assignment）. After that, these different features from the same image are fused by DSmT in order to gain the observation sequence of HMM. Then, the sequential information is fused by utilizing HMM to finish the automatic recognition of aircraft with varied multiple postures by calculating the similarities between the observation sequence and each HMM. Finally, according to the simulation experiment results, this algorithm has a high exact recognition rate even for aircraft with greatly varied postures. Simultaneously, this algorithm can also satisfy the requirement of aircraft target recognition in realtime. In addition, this algorithm can also guarantee a high recognition rate of aircraft target when the number T of consecutive occulted frames is less than or equal to six.

关 键 词：序列飞机 目标识别 多特征融合 DSmT推理 概率神经网络 序列信息融合 隐马尔可夫模型 

分 类 号：TP202[自动化与计算机技术—检测技术与自动化装置] TP391.41[自动化与计算机技术—控制科学与工程]