This work describes an end-to-end approach for real-time human action recognition from raw depth image-sequences. The proposal is based on a 3D fully convolutional neural network, named 3DFCNN, which automatically encodes spatio-temporal patterns from raw depth sequences. The described 3D-CNN allows actions classification from the spatial and temporal encoded information of depth sequences. The use of depth data ensures that action recognition is carried out protecting people’s privacy, since their identities can not be recognized from these data. The proposed 3DFCNN has been optimized to reach a good performance in terms of accuracy while working in real-time. Then, it has been evaluated and compared with other state-of-the-art systems in three widely used public datasets with different characteristics, demonstrating that 3DFCNN outperforms all the non-DNN-based state-of-the-art methods with a maximum accuracy of 83.6% and obtains results that are comparable to the DNN-based approaches, while maintaining a much lower computational cost of 1.09 seconds, what significantly increases its applicability in real-world environments.

This work proposes and compare two different approaches for real-time human action recognition (HAR) from raw depth video sequences. Both proposals are based on the convolutional long short-term memory unit, namely ConvLSTM, with differences in the architecture and the long-term learning. The former uses a video-length adaptive input data generator (stateless) whereas the latter explores thestatefulability of general recurrent neural networks but is applied in the particular case of HAR. This stateful property allows the model to accumulate discriminative patterns from previous frames without compromising computer memory. Furthermore, since the proposal uses only depth information, HAR is carried out preserving the privacy of people in the scene, since their identities can not be recognized. Both neural networks have been trained and tested using the large-scale NTU RGB+D dataset. Experimental results show that the proposed models achieve competitive recognition accuracies with lower computational cost compared with state-of-the-art methods and prove that, in the particular case of videos, the rarely-used stateful mode of recurrent neural networks significantly improves the accuracy obtained with the standard mode. The recognition accuracies obtained are 75.26% (CS) and 75.45% (CV) for the stateless model, with an average time consumption per video of 0.21 s, and 80.43% (CS) and 79.91%(CV) with 0.89 s for the stateful one.

Surveillance cameras are being installed in many primary daily living places to maintain public safety. In this video-surveillance context, anomalies occur only for a very short time, and very occasionally. Hence, manual monitoring of such anomalies may be exhaustive and monotonous, resulting in a decrease in reliability and speed in emergency situations due to monitor tiredness. Within this framework, the importance of automatic detection of anomalies is clear, and, therefore, an important amount of research works have been made lately in this topic. According to these earlier studies, supervised approaches perform better than unsupervised ones. However, supervised approaches demand manual annotation, making dependent the system reliability of the different situations used in the training (something difficult to set in anomaly context). In this work, it is proposed an approach for anomaly detection in video-surveillance scenes based on a weakly supervised learning algorithm. Spatio-temporal features are extracted from each surveillance video using a temporal convolutional 3D neural network (T-C3D). Then, a novel ranking loss function increases the distance between the classification scores of anomalous and normal videos, reducing the number of false negatives. The proposal has been evaluated and compared against state-of-art approaches, obtaining competitive performance without fine-tuning, which also validates its generalization capability. In this paper, the proposal design and reliability is presented and analyzed, as well as the aforementioned quantitative and qualitative evaluation in-the-wild scenarios, demonstrating its high sensitivity in anomaly detection in all of them.

In the analysis of biological time series, the state space is comprised of a framework for the study of systems with presumably deterministic and stationary properties. However, a physiological experiment typically captures an observable that characterizes the temporal response of the physiological system under study; the dynamic variables that make up the state of the system at any time are not available. Only from the acquired observations should state vectors be reconstructed to emulate the different states of the underlying system. This is what is known as the reconstruction of the state space, called the phase space in real-world signals, in many cases satisfactorily resolved using the method of delays. Each state vector consists of m components, extracted from successive observations delayed a time $\tau$ . The morphology of the geometric structure described by the state vectors, as well as their properties depends on the chosen parameters $\tau$ and m. The real dynamics of the system under study is subject to the correct determination of the parameters $\tau$ and m. Only in this way can be deduced features have true physical meaning, revealing aspects that reliably identify the dynamic complexity of the physiological system. The biological signal presented in this work, as a case study, is the photoplethysmographic (PPG) signal. We find that m is five for all the subjects analyzed and that $\tau$ depends on the time interval in which it is evaluated. The Hénon map and the Lorenz flow are used to facilitate a more intuitive understanding of the applied techniques.