The aim of the presented investigation is to explore the time gap between an anomaly appearance in continuously measured parameters of the device and a failure, related to the end of the remaining resource of the device-critical component. In this investigation, we propose a recurrent neural network to model the time series of the parameters of the healthy device to detect anomalies by comparing the predicted values with the ones actually measured. An experimental investigation was performed on SCADA estimates received from different wind turbines with failures. A recurrent neural network was used to predict the temperature of the gearbox. The comparison of the predicted temperature values and the actual measured ones showed that anomalies in the gearbox temperature could be detected up to 37 days before the failure of the device-critical component. The performed investigation compared different models that can be used for temperature time-series modeling and the influence of selected input features on the performance of temperature anomaly detection.

The presented research faces the problem of early detection of sepsis for patients in the Intensive Care Unit. The PhysioNet/Computing in Cardiology Challenge 2019 facilitated the development of automated, open-source algorithms for the early detection of sepsis from clinical data. A labeled clinical records dataset for training and verification of the algorithms was provided by the challenge organizers. However, a relatively small number of records with sepsis, supported by Sepsis-3 clinical criteria, led to highly unbalanced dataset (only 2% records with sepsis label). A high number of unbalanced data records is a great challenge for machine learning model training and is not suitable for training classical classifiers. To address these issues, a method taking into the account the amount of time the patients spent in the intensive care unit (ICU) was proposed. The proposed method uses two separate ensemble models, one trained on patient records under 56 h in the ICU, and another for patients who stayed longer than 56 h. A solution including feature selection and weighting based training on imbalanced data was proposed in this paper. In addition, several performance metrics were investigated. Results show, that for successful prediction, a particular model having few or more predictors based on the length of stay in the Intensive Care Unit should be applied.

Creation of head 3D models from videos or pictures of the head by using close-range photogrammetry techniques has many applications in clinical, commercial, industrial, artistic, and entertainment areas. This work aims to create a methodology for improving 3D head reconstruction, with a focus on using selfie videos as the data source. Then, using this methodology, we seek to propose changes for the general-purpose 3D reconstruction algorithm to improve the head reconstruction process. We define the improvement of the 3D head reconstruction as an increase of reconstruction quality (which is lowering reconstruction errors of the head and amount of semantic noise) and reduction of computational load. We proposed algorithm improvements that increase reconstruction quality by removing image backgrounds and by selecting diverse and high-quality frames. Algorithm modifications were evaluated on videos of the mannequin head. Evaluation results show that baseline reconstruction is improved 12 times due to the reduction of semantic noise and reconstruction errors of the head. The reduction of computational demand was achieved by reducing the frame number needed to process, reducing the number of image matches required to perform, reducing an average number of feature points in images, and still being able to provide the highest precision of the head reconstruction.

Pedestrian detection is an essential task for computer vision and the automotive industry. Complex systems like advanced driver-assistance systems are based on far-infrared data sensors, used to detect pedestrians at nighttime, fog, rain, and direct sun situations. The robust pedestrian detector should work in severe weather conditions. However, only a few datasets include some examples of far-infrared images with distortions caused by atmospheric precipitation and dirt covering sensor optics. This paper proposes the deep learning-based data augmentation technique to enrich far-infrared images collected in good weather conditions by distortions, similar to those caused by bad weather. The six most accurate and fast detectors (TinyV3, TinyL3, You Only Look Once (YOLO)v3, YOLOv4, ResNet50, and ResNext50), performing faster than 15 FPS, were trained on 207,001 annotations and tested on 156,345 annotations, not used for training. The proposed data augmentation technique showed up to a 9.38 mean Average Precision (mAP) increase of pedestrian detection with a maximum of 87.02 mAP (YOLOv4). Proposed in this paper detectors’ Head modifications based on a confidence heat-map gave an additional boost of precision for all six detectors. The most accurate current detector, based on YOLOv4, reached up to 87.20 mAP during our experimental tests.

A typical pedestrian protection system requires sophisticated hardware and robust detection algorithms. To solve these problems the existing systems use hybrid sensors where mono and stereo vision merged with active sensors. One of the most assuring pedestrian detection sensors is far infrared range camera. The classical pedestrian detection approach based on Histogram of oriented gradients is not robust enough to be applied in devices which consumers can trust. An application of deep neural network-based approach is able to perform with significantly higher accuracy. However, the deep learning approach requires a high number of labeled data examples. The investigation presented in this paper aimed the acceleration of pedestrian labeling in far-infrared image sequences. In order to accelerate pedestrian labeling in far-infrared camera videos, we have integrated the YOLOv3 object detector into labeling software. The verification of the pre-labeled results was around eleven times faster than manual labeling of every single frame.

This paper describes the hardware implementation of the modified speeded-up robust features (SURF) algorithm. Field Programmable Gate Array (FPGA) is selected as a hardware platform for implementation of the modified SURF algorithm. FPGA base platform is selected to ensure features extraction in real-time video. A sliding window buffer is used to store shifted pixels of integral video frame at each clock signal. The buffer is shared with Hessian determinant, orientation and descriptor units. The interest points are searched in 8 scales in parallel by applying the non-maximal suppression. The orientation vector and simplified descriptor is computed for each feature in 6 scales. The proposed architecture is implemented using VHDL and achieves real-time orientation and descriptor calculation on 60 fps 640×480 video stream only on 25 MHz clock.