The accurate classification of activity patterns based on radar signatures is still an open problem and is key to detect anomalous behavior for security and health applications. This paper presents a novel iterative convolutional neural networks strategy with an autocorrelation pre-processing instead of the traditional micro-Doppler image pre-processing to classify activities or subjects accurately. The proposed strategy uses an iterative deep learning framework for the automatic definition and extraction of features. This is followed by a traditional supervised learning classifier to label the different activities. Using three human subjects and their real motion captured data, twelve thousand radar signatures were simulated by varying additive white Gaussian noise. Additionally, 6720 experimental radar signatures were captured with a frequency-modulated continuous radar at 5.8GHz with 400MHz of instantaneous bandwidth from seven activities using one subject and 4800 signatures from five subjects while walking. The simulated and experimental data were both used to validate our proposed method. With SNR varying from –20 to 20dB with 88.74% average accuracy at –10dB and 100% peak accuracy at 15dB. The proposed Iterative Convolutional Neural Networks followed with Random Forests (ICNNRF) does not only outperform the feature-based methods using micro-Doppler images but also the classification methods using other types of supervised classifiers after our proposed iterative convolutional neural network.
Classification of different human activities using micro-Doppler data and features is considered in this study, focusing on the distinction between walking and running. 240 recordings from 2 different human subjects were collected in a series of simulations performed in the real motion data from the Carnegie Mellon University Motion Capture Database. The maximum the micro-Doppler frequency shift and the period duration are utilized as two classification criterions. Numerical results are compared against several classification techniques including the Linear Discriminant Analysis (LDA), Naïve Bayes (NB), K-nearest neighbors (KNN), Support Vector Machine(SVM) algorithms. The performance of different classifiers is discussed aiming at identifying the most appropriate features for the walking and running classification.
This study introduces an innovative approach by incorporating statistical offset features, range profiles, time–frequency analyses, and azimuth–range–time characteristics to effectively identify various human daily activities. Our technique utilizes nine feature vectors consisting of six statistical offset features and three principal component analysis network (PCANet) fusion attributes. These statistical offset features are derived from combined elevation and azimuth data, considering their spatial angle relationships. The fusion attributes are generated through concurrent 1D networks using CNN-BiLSTM. The process begins with the temporal fusion of 3D range–azimuth–time data, followed by PCANet integration. Subsequently, a conventional classification model is employed to categorize a range of actions. Our methodology was tested with 21,000 samples across fourteen categories of human daily activities, demonstrating the effectiveness of our proposed solution. The experimental outcomes highlight the superior robustness of our method, particularly when using the Margenau–Hill Spectrogram for time–frequency analysis. When employing a random forest classifier, our approach outperformed other classifiers in terms of classification efficacy, achieving an average sensitivity, precision, F1, specificity, and accuracy of 98.25%, 98.25%, 98.25%, 99.87%, and 99.75%, respectively.
This paper presents a novel approach that applies WiFi-based IQ data and time–frequency images to classify human activities automatically and accurately. The proposed strategy first uses the Choi–Williams distribution transform and the Margenau–Hill spectrogram transform to obtain the time–frequency images, followed by the offset and principal component analysis (PCA) feature extraction. The offset features were extracted from the IQ data and several spectra with maximum energy values in the time domain, and the PCA features were extracted via the whole images and several image slices on them with rich unit information. Finally, a traditional supervised learning classifier was used to label various activities. With twelve-thousand experimental samples from four categories of WiFi signals, the experimental data validated our proposed method. The results showed that our method was more robust to varying image slices or PCA numbers over the measured dataset. Our method with the random forest (RF) classifier surpassed the method with alternative classifiers on classification performance and finally obtained a 91.78% average sensitivity, 91.74% average precision, 91.73% average F1-score, 97.26% average specificity, and 95.89% average accuracy.
This paper proposes a tri-modal speech reconstruction method that synergistically fuses laser speckle vibrometry, depth-camera articulatory capture, and ultra-wideband (UWB) radar. By exploiting the RVTALL multimodal dataset, the method extracts highly complementary physical information: laser signals capture physiological vibrations homologous to voice sources, depth cameras record articulatory geometry, and UWB radar detects radio-frequency micro-motions. To maximize the utility of these heterogeneous data streams, the framework adopts a two-stage training paradigm, where modality-specific encoders are pretrained with an auxiliary classification objective and then frozen for label-free log-mel spectrogram regression. A hybrid loss combining global spectral consistency and local dynamic constraints is designed to optimize fine-grained formant transitions and energy onsets. The proposed method demonstrates significant superiority over single-modality approaches without requiring any textual or phonetic supervision. Experimental results under a strict leave-two-subjects-out protocol show that the full tri-modal system achieves an 81.7% word-level Whisper accuracy, a 16.3% absolute improvement over the single laser modality (65.4%). Furthermore, the wav2vec 2.0 sentence-level cosine similarity reaches 0.932 (compared to 0.799 for laser only), and the word and sentence mean opinion scores (MOS) reach 3.08 and 2.67, respectively. These results confirm that the rich physical complementarity within the tri-modal dataset enables highly intelligible, natural, and noise-resilient speech recovery.
This paper presents a general approach for extracting the specific feature directly from all the Intrinsic Mode Functions (IMF). It is believed that every IMF is possessed with its own physical meaning. However, a detailed signature is a part of the specially determined IMF. Due to various IMF numbers as well as IMF frequency mean deviations, it costs much time to choose the certain IMF, which is also easy to make mistakes. In this study, in order to prevent from making the mistake of frequency mean deviations, a novel algorithm is put forward on the basis of the analysis between adjacent IMFs with the support of the mathematical statistics method. According to the numerical results, the proposed algorithm is in a good conformity to the simulation one.