In this paper, an accurate approach for vehicle counting in videos using Mask R-CNN and KLT tracker is proposed. Vehicle detection is performed for each N frames using Mask R-CNN instance segmentation model. This model outperforms other deep learning models that using bounding box detection as it provides a segmentation mask for each detected object, the outperformance comes up clearly in cases of occlusions. Once the objects are detected, their corner points are extracted and tracked. An efficient method is introduced to assign point trajectories to their corresponding detected vehicles. The proposed counting algorithm distinguishes precisely between the new vehicles and the counted ones. The experiments performed on diverse challenging videos show excellent results compared to state-of-the-art counting methods.

A trained neural network classifier is often used to detect cardiac problems by the classification of heart sound signals, also known as phonocardiogram (PCG) signals. The choice of an appropriate training optimization algorithm for such a classification problem, on the other hand, is still being debated. In this paper, we use the bidirectional long short-term memory (biLSTM) network for the classification of sequences of short-time features extracted from labelled PCG signals. The classification performance of four different trained biLSTM models is described in terms of three different optimization algorithms that are used to train the classifier. The elaborated results on testing PCG signals show that the biLSTM classifier performs better when trained with the stochastic gradient descent with momentum (SGDM) algorithm than when trained with the RMSprop (root mean squared propagation) optimizer or the adaptive moment (ADAM) optimization algorithm. Furthermore, this classification method outperforms a baseline method.

In Gaussian model-based multichannel audio source separation, the likelihood of observed mixtures of source signals is parametrized by source spectral variances and by associated spatial covariance matrices. These parameters are estimated by maximizing the likelihood through an expectation-maximization algorithm and used to separate the signals by means of multichannel Wiener filtering. We propose to estimate these parameters by applying nonnegative factorization based on prior information on source variances. In the nonnegative factorization, spectral basis matrices can be defined as the prior information. The matrices can be either extracted or indirectly made available through a redundant library that is trained in advance. In a separate step, applying nonnegative tensor factorization, two algorithms are proposed in order to either extract or detect the basis matrices that best represent the power spectra of the source signals in the observed mixtures. The factorization is achieved by minimizing the β-divergence through multiplicative update rules. The sparsity of factorization can be controlled by tuning the value of β. Experiments show that sparsity, rather than the value assigned to β in the training, is crucial in order to increase the separation performance. The proposed method was evaluated in several mixing conditions. It provides better separation quality with respect to other comparable algorithms.

Multichannel sparse representation of acoustic sources has shown to provide an attractive framework for source separation. The multichannel sparse modeling assumes an ability to describe signals as linear combinations of few atoms from a pre-specified dictionary. The dictionary is built by simulating room impulse responses on a grid of locations, exploiting a prior knowledge on the room geometry and reflection coefficients. However, due to the simplified modeling, any mismatch between the simulated and true observed RIRs would generate a considerable distortion in the recovered output signals. In this work we propose an unsupervised adaptation of the dictionary through a semi-blind weighted Natural Gradient, assuming spatio-temporal source sparseness. The system continuously adapts the atoms with the incoming data, improving the match between the dictionary and the true mixing parameters. Results over simulated data show that the proposed framework is a promising solution to underdetermined convolutive source separation in difficult acoustic scenarios.

Artificial neural networks have proven high efficiency in the classification of features in many application fields. It is widely proved that dynamic recurrent networks perform better than any other types of networks, for the feedback connections and the embedded memory. In this paper, we use a dynamic recurrent network called the nonlinear autoregressive network with exogenous inputs (NARX) for the normal/abnormal classification of phonocardiography (PCG) signals. Up to the authors' knowledge, this is the first time that the NARX network is used for such classification task. The network is trained and tested using a dataset of 6316 feature vectors, each vector of length 27 features. The evaluation of this network is achieved at different sizes of the embedded memory. The experimental results show that the NARX network outperforms other two networks, and its performance increases as a function of the memory size till a certain value. Further increase of the memory size degrades the performance as a results of model overfitting.