MediNet: transfer learning approach with MediNet medical visual database.

The rapid development of machine learning has increased interest in the use of deep learning methods in medical research. Deep learning in the medical field is used in disease detection and classification problems in the clinical decision-making process. Large amounts of labeled datasets are often required to train deep neural networks; however, in the medical field, the lack of a sufficient number of images in datasets and the difficulties encountered during data collection are among the main problems. In this study, we propose MediNet, a new 10-class visual dataset consisting of Rontgen (X-ray), Computed Tomography (CT), Magnetic Resonance Imaging (MRI), Ultrasound, and Histopathological images such as calcaneal normal, calcaneal tumor, colon benign colon adenocarcinoma, brain normal, brain tumor, breast benign, breast malignant, chest normal, chest pneumonia. AlexNet, VGG19-BN, Inception V3, DenseNet 121, ResNet 101, EfficientNet B0, Nested-LSTM + CNN, and proposed RdiNet deep learning algorithms are used in the transfer learning for pre-training and classification application. Transfer learning aims to apply previously learned knowledge in a new task. Seven algorithms were trained with the MediNet dataset, and the models obtained from these algorithms, namely feature vectors, were recorded. Pre-training models were used for classification studies on chest X-ray images, diabetic retinopathy, and Covid-19 datasets with the transfer learning technique. In performance measurement, an accuracy of 94.84% was obtained in the traditional classification study for the InceptionV3 model in the classification study performed on the Chest X-Ray Images dataset, and the accuracy was increased 98.71% after the transfer learning technique was applied. In the Covid-19 dataset, the classification success of the DenseNet121 model before pre-trained was 88%, while the performance after the transfer application with MediNet was 92%. In the Diabetic retinopathy dataset, the classification success of the Nested-LSTM + CNN model before pre-trained was 79.35%, while the classification success was 81.52% after the transfer application with MediNet. The comparison of results obtained from experimental studies observed that the proposed method produced more successful results.

Pedestrian Origin-Destination Estimation Based on Multi-Camera Person Re-Identification.

Pedestrian origin-destination (O-D) estimates that record traffic flows between origins and destinations, are essential for the management of pedestrian facilities including pedestrian flow simulation in the planning phase and crowd control in the operation phase. However, current O-D data collection techniques such as surveys, mobile sensing using GPS, Wi-Fi, and Bluetooth, and smart card data have the disadvantage that they are either time consuming and costly, or cannot provide complete O-D information for pedestrian facilities without entrances and exits or pedestrian flow inside the facilities. Due to the full coverage of CCTV cameras and the huge potential of image processing techniques, we address the challenges of pedestrian O-D estimation and propose an image-based O-D estimation framework. By identifying the same person in disjoint camera views, the O-D trajectory of each identity can be accurately generated. Then, state-of-the-art deep neural networks (DNNs) for person re-ID at different congestion levels were compared and improved. Finally, an O-D matrix based on trajectories was generated and the resident time was calculated, which provides recommendations for pedestrian facility improvement. The factors that affect the accuracy of the framework are discussed in this paper, which we believe could provide new insights and stimulate further research into the application of the Internet of cameras to intelligent transport infrastructure management.

InSiNet: a deep convolutional approach to skin cancer detection and segmentation.

Cancer is among the common causes of death around the world. Skin cancer is one of the most lethal types of cancer. Early diagnosis and treatment are vital in skin cancer. In addition to traditional methods, method such as deep learning is frequently used to diagnose and classify the disease. Expert experience plays a major role in diagnosing skin cancer. Therefore, for more reliable results in the diagnosis of skin lesions, deep learning algorithms can help in the correct diagnosis. In this study, we propose InSiNet, a deep learning-based convolutional neural network to detect benign and malignant lesions. The performance of the method is tested on International Skin Imaging Collaboration HAM10000 images (ISIC 2018), ISIC 2019, and ISIC 2020, under the same conditions. The computation time and accuracy comparison analysis was performed between the proposed algorithm and other machine learning techniques (GoogleNet, DenseNet-201, ResNet152V2, EfficientNetB0, RBF-support vector machine, logistic regression, and random forest). The results show that the developed InSiNet architecture outperforms the other methods achieving an accuracy of 94.59%, 91.89%, and 90.54% in ISIC 2018, 2019, and 2020 datasets, respectively. Since the deep learning algorithms eliminate the human factor during diagnosis, they can give reliable results in addition to traditional methods.

A Recurrent Deep Network for Estimating the Pose of Real Indoor Images from Synthetic Image Sequences.

Recently, deep convolutional neural networks (CNN) have become popular for indoor visual localisation, where the networks learn to regress the camera pose from images directly. However, these approaches perform a 3D image-based reconstruction of the indoor spaces beforehand to determine camera poses, which is a challenge for large indoor spaces. Synthetic images derived from 3D indoor models have been used to eliminate the requirement of 3D reconstruction. A limitation of the approach is the low accuracy that occurs as a result of estimating the pose of each image frame independently. In this article, a visual localisation approach is proposed that exploits the spatio-temporal information from synthetic image sequences to improve localisation accuracy. A deep Bayesian recurrent CNN is fine-tuned using synthetic image sequences obtained from a building information model (BIM) to regress the pose of real image sequences. The results of the experiments indicate that the proposed approach estimates a smoother trajectory with smaller inter-frame error as compared to existing methods. The achievable accuracy with the proposed approach is 1.6 m, which is an improvement of approximately thirty per cent compared to the existing approaches. A Keras implementation can be found in our Github repository.

A Robust and Adaptive Complementary Kalman Filter Based on Mahalanobis Distance for Ultra Wideband/Inertial Measurement Unit Fusion Positioning.

Ultra wideband (UWB) has been a popular technology for indoor positioning due to its high accuracy. However, in many indoor application scenarios UWB measurements are influenced by outliers under non-line of sight (NLOS) conditions. To detect and eliminate outlying UWB observations, we propose a UWB/Inertial Measurement Unit (UWB/IMU) fusion filter based on a Complementary Kalman Filter to track the errors of position, velocity and direction. By using the least squares method, the positioning residual of the UWB observation is calculated, the robustness factor of the observation is determined, and an observation weight is dynamically set. When the robustness factor does not exceed a pre-defined threshold, the observed value is considered trusted, and adaptive filtering is used to track the system state, while the abnormity of system state, which might be caused by IMU data exceptions or unreasonable noise settings, is detected by using Mahalanobis distance from the observation to the prior distribution. When the robustness factor exceeds the threshold, the observed value is considered abnormal, and robust filtering is used, whereby the impact of UWB data exceptions on the positioning results is reduced by exploiting Mahalanobis distance. Experimental results show that the observation error can be effectively estimated, and the proposed algorithm can achieve an improved positioning accuracy when affected by outlying system states of different quantity as well as outlying observations of different proportion.

Sensors for Indoor Mapping and Navigation.

With the growth of cities and increased urban population there is a growing demand for spatial information of large indoor environments.[...].

User-Independent Motion State Recognition Using Smartphone Sensors.

The recognition of locomotion activities (e.g., walking, running, still) is important for a wide range of applications like indoor positioning, navigation, location-based services, and health monitoring. Recently, there has been a growing interest in activity recognition using accelerometer data. However, when utilizing only acceleration-based features, it is difficult to differentiate varying vertical motion states from horizontal motion states especially when conducting user-independent classification. In this paper, we also make use of the newly emerging barometer built in modern smartphones, and propose a novel feature called pressure derivative from the barometer readings for user motion state recognition, which is proven to be effective for distinguishing vertical motion states and does not depend on specific users' data. Seven types of motion states are defined and six commonly-used classifiers are compared. In addition, we utilize the motion state history and the characteristics of people's motion to improve the classification accuracies of those classifiers. Experimental results show that by using the historical information and human's motion characteristics, we can achieve user-independent motion state classification with an accuracy of up to 90.7%. In addition, we analyze the influence of the window size and smartphone pose on the accuracy.

3D modeling of building indoor spaces and closed doors from imagery and point clouds.

3D models of indoor environments are increasingly gaining importance due to the wide range of applications to which they can be subjected: from redesign and visualization to monitoring and simulation. These models usually exist only for newly constructed buildings; therefore, the development of automatic approaches for reconstructing 3D indoors from imagery and/or point clouds can make the process easier, faster and cheaper. Among the constructive elements defining a building interior, doors are very common elements and their detection can be very useful either for knowing the environment structure, to perform an efficient navigation or to plan appropriate evacuation routes. The fact that doors are topologically connected to walls by being coplanar, together with the unavoidable presence of clutter and occlusions indoors, increases the inherent complexity of the automation of the recognition process. In this work, we present a pipeline of techniques used for the reconstruction and interpretation of building interiors based on point clouds and images. The methodology analyses the visibility problem of indoor environments and goes in depth with door candidate detection. The presented approach is tested in real data sets showing its potential with a high door detection rate and applicability for robust and efficient envelope reconstruction.

Accuracy and resolution of Kinect depth data for indoor mapping applications.

Consumer-grade range cameras such as the Kinect sensor have the potential to be used in mapping applications where accuracy requirements are less strict. To realize this potential insight into the geometric quality of the data acquired by the sensor is essential. In this paper we discuss the calibration of the Kinect sensor, and provide an analysis of the accuracy and resolution of its depth data. Based on a mathematical model of depth measurement from disparity a theoretical error analysis is presented, which provides an insight into the factors influencing the accuracy of the data. Experimental results show that the random error of depth measurement increases with increasing distance to the sensor, and ranges from a few millimeters up to about 4 cm at the maximum range of the sensor. The quality of the data is also found to be influenced by the low resolution of the depth measurements.

Assessment of relative accuracy of AHN-2 laser scanning data using planar features.

AHN-2 is the second part of the Actueel Hoogtebestand Nederland project, which concerns the acquisition of high-resolution altimetry data over the entire Netherlands using airborne laser scanning. The accuracy assessment of laser altimetry data usually relies on comparing corresponding tie elements, often points or lines, in the overlapping strips. This paper proposes a new approach to strip adjustment and accuracy assessment of AHN-2 data by using planar features. In the proposed approach a transformation is estimated between two overlapping strips by minimizing the distances between points in one strip and their corresponding planes in the other. The planes and the corresponding points are extracted in an automated segmentation process. The point-to-plane distances are used as observables in an estimation model, whereby the parameters of a transformation between the two strips and their associated quality measures are estimated. We demonstrate the performance of the method for the accuracy assessment of the AHN-2 dataset over Zeeland province of The Netherlands. The results show vertical offsets of up to 4 cm between the overlapping strips, and horizontal offsets ranging from 2 cm to 34 cm.