基于CK-Hough联合算法的人体微多普勒频率估计

Human micro-Doppler frequency estimation by CK-Hough joint algorithm

为了准确地从雷达回波信号中提取运动目标特定部位的微多普勒频率,本文提出一种新颖的CKHough算法,该算法有效地结合了聚类分析和K近邻-霍夫(KNN-Hough)算法。首先,通过短时傅里叶变换获取雷达回波信号的时频谱图;其次,利用自适应模糊C均值算法对时频图进行聚类分析,在这一过程中,本文采用数据预处理技术自适应调整聚类类别数c以适应多样化应用场景,从而获得人体各散射部位的频域范围,有效地抑制了分量间的相互干扰;第三,通过改进度量函数的K近邻算法增强相邻时刻聚类结果的相关性,拟合各部位的瞬时频率曲线;最后,采用霍夫变换动态调整度量函数中权值μ的取值,得到目标微多普勒频率的精确估计结果。研究结果表明:本文提出的CK-Hough提取了直/曲线行走场景下人类目标四肢的微多普勒频率;与传统的峰值搜索算法、线性预测维特比算法以及基于Bezier-Hough模型的频率拟合算法相比,本文提出的CK-Hough算法在直线行走实验场景下,总频率的估计误差率分别降低了40.40%、45.47%和26.16%;在曲线行走实验场景下,其估计误差率分别降低了58.35%、68.35%和41.65%。

To accurately extract the micro-Doppler(m-D)frequency of a specific part of a moving target from the radar echo signal,a novel CK-Hough algorithm was proposed,which effectively combines the cluster analysis and the K Nearest Neighbor-Hough(KNN-Hough)algorithm.Firstly,the time-frequency spectrogram of the echo signal was obtained by a short-time Fourier transform(STFT).Secondly,the time-frequency spectrogram was clustered by the adaptive fuzzy C-means(AFCM)algorithm to obtain the frequency ranges of the different scattering parts of the human body and effectively inhibit the mutual interference between the components.The data pre-processing technique was used to adaptively adjust the number of clustering categories c,which adapts to diversified application scenarios.Thirdly,the improved KNN algorithm was utilized to enhance the correlation between the clustering results of adjacent moments and to fit the instantaneous frequency curve.Finally,the Hough transform was employed to determine the weight value ofμ.The results show that the CK-Hough proposed in the paper extracts the m-D frequencies of human limbs in straight/curved walking scenarios.Compared with the traditional peak search algorithm,linear prediction Viterbi algorithm and frequency fitting algorithm based on the Bezier-Hough model,the CK-Hough algorithm proposed reduces the estimation error rate of frequency by 40.40%,45.47%and 26.16%,respectively.Furthermore,in the curve walking experimental scenario,its estimation error rate decreases by 58.35%,68.35%and 41.65%,respectively.