Chundi Zheng

In order to achieve automatic behavioral monitoring of farming animals, this paper proposes a convolutional neural network-based behavioral recognition method for lactating sows. The behavioral data streams of lactating sows are collected by wearable sensors embedded with an 3D-accelerometer and a 3D-gyroscope and used to recognize six types of behaviors, including movement, drinking, eating, nursing, sleeping and lying. Among these behaviors, nursing, lying and sleeping can be classified as similar static behaviors. Based on the action images constructed with sensor data streams, convolutional neural networks is leveraged for the purpose of distinguishing static behaviors of the lactating sow. To address the problem of insufficient training data, we use the data augmentation technique. Experimental results verify the data augmentation method’s effectiveness and show that our proposed behavioral monitoring method has greater advantages in terms of accuracy than traditional machine learning methods. The research results have implications for behavioral monitoring and health assessment of lactating sows.

High-resolution and low-complexity Direction Of Arrival (DOA) estimation deserves a warm welcome in a real radar sensor. However, achieving this goal is nontrivial. For the DOA estimation of the uniform linear array, we offer an efficient approach to reach a good compromise between complexity, resolution and accuracy. Under the covariance fitting framework, we construct a sparse Non-Negative Least Squares (NNLS) that is solved by the Alternating Direction Method of Multipliers (ADMM) at a low computational complexity. The regularization item involved in the ADMM procedure yields an inherent improvement on robustness. Moreover, a large-scale matrix inversion that is a computationally expensive operation in the ADMM procedure is recursively computed by exploiting the diagonal matrix property of the regularization item, resulting in a reduction on complexity. In addition, the pre-estimation of the noise variance further reduces the scale of the sparse NNLS problem and alleviates the ill-conditioning effect to some extent. Compared to several sparse covariance fitting methods, both the simulation data and the 77 GHz radar measured data demonstrate that the proposed algorithm has the good resolution, robustness, and accuracy performance with lower computationally cost.

Achieving high-accuracy and non-contact heart rate monitoring via radar is a non-trivial task under the interference from respiration and its harmonics. In order to eliminate the interference, we propose a weighted reconstruction method of the second harmonic signal of the heartbeat for the heart rate estimation. The proposed method combines the Variational Modal Decomposition (VMD) algorithm with the prior information about the second harmonic of the heartbeat. Namely, after a sophisticated endeavor for preprocessing, the chest wall displacement signal is decomposed into several quasi-orthogonal components by VMD and then those components whose spectrum are within the second harmonic frequency band of the heartbeat are selected to reconstruct the second harmonic signal of the heartbeat. For highlighting more likely second harmonic components and suppressing undesired noise components, the energy of each selected component is used as a weight to adjust its proportion in the reconstructed signal. Consequently, the heart rate estimation is achieved indirectly by estimating the spectral peak of the second harmonic signal. The proposed method not only avoids the strong interference from the respiratory signal and its higher harmonics on heart rate estimation, but also predigests the parameter setting in the spectral peak estimation model. Experimental results from 15 subjects demonstrate that the proposed method can achieve higher measurement accuracy compared with other several methods based on a 77 GHz FMCW radar.

Using wideband signals to improve the resolution performance of forward-looking imaging radar sensors is a very valuable topic. However, the frequency point and angle information are coupled in the array manifold for the wideband array signal, making the high-resolution estimation of range and angle extremely difficult under low complexity. In this paper, we propose a novel algorithm for the wideband forward-looking imaging radar. Firstly, the time delays are estimated by the estimation of signal parameters via rotation invariance techniques (ESPRIT). Then, multiple sets of filters are constructed by using the estimated delays and frequency information. The filters are equipped with sliding window technology to tackle the angle-frequency coupling problem encountered by wideband radar. Through these filters, the received signals are divided into several clusters that correspond to different time delays. Next, the relaxation (RELAX) algorithm is used to estimate the angle of the wideband signal in each cluster. Finally, the range refinement is carried out by using data from all antennas. Compared with the existing algorithms, the proposed Hybrid ESPRIT and RELAX algorithm with clustEring (HERE) decouples the frequency point and angle information for wideband forward-looking imaging radar and then achieves better imaging performance. A series of numerical experiments, including simulation experiments and actual data collected from a wideband forward-looking imaging prototype, have verified the superiority of the algorithm.

Combining the covariance matching criterion with sparse representation, much effort was devoted to improve the angular resolution. However, the requirement of hyperparameters or the high sidelobe makes them unsatisfactory in practical radar applications. In this paper, we propose a Sparse Capon-like Weighted Quadratic Estimator (SCWQE) with hyperparameter-free and apply it to a 77GHz Frequency Modulated Continuous Wave (FMCW) radar for the high-resolution Direction Of Arrival (DOA) estimation. Under the assumption of the uncorrelated sources, SCWQE is formulated by converting the covariance matching criterion to a quadratic function with respect to the unknown power spectrum, in which a symmetrical and non-diagonal matrix called as the Capon-like spectrum is embedded into the quadratic form. The consequent multiple weighting operation would further promote the sparsity because multiple and different weighting values are exerted on each element of the spatial power spectrum. This is fundamentally different from the traditional weighted approach that employs the diagonal weighting matrix and assigns a single weighting value to each element. Therefore, SCWQE could further eliminate the spurious peaks and sharpen the desired peaks. Numerical examples including simulations and actual data collected from a 77GHz FMCW radar sensor demonstrate that, the proposed SCWQE algorithm produces better accuracy of DOA estimation and higher angular resolution compared to some typical sparse recovery methods.

The power of end-to-end deep learning techniques to automatically learn latent high-level features from raw signals has been demonstrated in numerous application fields, however, few studies systematically investigate how to properly encode the time-series firings of binary environment sensors that typically work in an event-triggering scheme and have irregular sampling rates for in-home human activity recognition. To this end, we here propose two different methods to process the streaming sensor readings and accordingly evaluate their combinations with deep learning models. Specifically, we divide the multichannel sensor events into segments and encode each segment into either a vector or two-dimensional matrix. Particularly, three different feature representations are presented for the vector form. Afterwards, we combine the encoded features with four typical deep learning models and optimize corresponding activity recognizers to study their sensitivity to different feature encodings. Furthermore, we include seven commonly used shallow classification models for comparison purposes. Finally, we conduct extensive experiments on three publicly available smart home datasets. Results indicate that the performance of both deep learning and shallow models is closely associated with the raw signal encodings and demonstrate the superiority of one-dimensional convolutional neural networks over its competitors in terms of generalization across scenarios. Besides, we preliminarily explore the influence of NULL class on an activity recognizer and experimentally show its negative impact on overall accuracy, enlightening relevant studies to consider it in developing a practical activity recognition system.

Pervasive computing greatly advances the automatic recognition and understanding of human activities and effectively bridges the gap between the low-level sensor signals and high-level human-centric applications. The inherent complexity of human behavior, however, inevitably poses a huge challenge to the design of a robust activity recognizer, especially in classifying similar activities. In this study, we present a hierarchical framework, named HierHAR, that infers on-going activities in a multi-stage process for better distinguishing similar activities and improving the overall performance. Specifically, we propose a data-driven approach, rather than heavily rely on prior domain knowledge, to automatically determining the relationships among activities. Afterwards, we use the relationships to organize the activities into a tree structure and accordingly construct and train a tree-based activity recognition model. Furthermore, we train a graph-based model that aims to reduce the compounding errors induced by the prediction process of the tree-based model. Finally, extensive comparative experiments are conducted on public datasets and results demonstrate the power of HierHAR in facilitating the automatic organization of activities and the design of hierarchical recognizers without prior knowledge about activities. Besides, the graph-based activity recognizer generally generalizes better across different scenarios and outperforms the tree-based model.

Traffic congestion reduces the working efficiency of taxis, thus lowering the income of taxi drivers. In most cases, taxi drivers will choose the least time-consuming route, which is usually the shortest in distance to the destination. However, during rush hours when traffic flow is heavy, the shortest route may not be the most time-efficient. A detour sometimes can save time for drivers but that may highly increase the fare for passengers. In order to optimize the balance between fare and profit, this paper proposes a framework, called ProfitMax, which will recommend a fixed price and a more profitable route by assigning the most suitable taxi for online taxi-hailing. The taxi's profit of completing an order is calculated by the income per minute. We use drivers' personal driving habit to estimate the basic cost of an order and then recommend the most profitable route. ProfitMax trains the personal fuel-consumption model and estimates the fuel consumption on drivers' smart devices. We use a real taxiing data set, including one-month taxis' GPS trajectories and car-OBD (on-board-diagnose) readings, for the performance evaluation. Experimental results show that ProfitMax estimated the fuel consumption more accurate than baselines and can also save more than 10% fuel.