In this paper, we propose a deep neural network-based method for estimating speed of vehicles on roads automatically from videos recorded using unmanned aerial vehicle (UAV). The proposed method includes the following; (1) detecting and tracking vehicles by analyzing the videos, (2) calculating the image scales using the distances between lanes on the roads, and (3) estimating the speeds of vehicles on the roads. Our method can automatically measure the speed of the vehicles from the only videos recorded using UAV without additional information in both directions on the roads simultaneously. In our experiments, we evaluate the performance of the proposed method with the visual data at four different locations. The proposed method shows 97.6% recall rate and 94.7% precision rate in detecting vehicles, and it shows error (root mean squared error) of 5.27 km/h in estimating the speeds of vehicles.

In this paper, we proposed a gait type classification method based on deep learning using a smart insole with various sensor arrays. We measured gait data using a pressure sensor array, an acceleration sensor array, and a gyro sensor array built into a smart insole. Features of gait pattern were then extracted using a deep convolution neural network (DCNN). In order to accomplish this, measurement data of continuous gait cycle were divided into unit steps. Pre-processing of data were then performed to remove noise followed by data normalization. A feature map was then extracted by constructing an independent DCNN for data obtained from each sensor array. Each of the feature maps was then combined to form a fully connected network for gait type classification. Experimental results for seven types of gait (walking, fast walking, running, stair climbing, stair descending, hill climbing, and hill descending) showed that the proposed method provided a high classification rate of more than 90%.

In this letter, we propose a novel method for classifying ambulatory activities using eight plantar pressure sensors within smart shoes. Using these sensors, pressure data of participants can be collected regarding level walking, stair descent, and stair ascent. Analyzing patterns of the ambulatory activities, we present new features with which to describe the ambulatory activities. After selecting critical features, a multi-class support vector machine algorithm is applied to classify these activities. Applying the proposed method to the experimental database, we obtain recognition rates up to 95.2% after six steps.

This paper proposes a whole new face image-based eye gaze estimation network to solve low generalization performance. Due to the high variance of facial appearance and environmental conditions, conventional methods in gaze estimation have low generalization performance and are easily overfitted to training subjects. To solve this problem, we adopt a self-attention mechanism that has better generalization performance. Nevertheless, applying self-attention directly to an image incurs a high computational cost. Thus, we introduce a new projection that uses convolution in the entire face image to accurate model the local context and reduce the computational cost of self-attention. The pro-posed model also includes deconvolution that transforms the down-sampled global context to the same size as the input so that spatial information is not lost. We confirmed through observations that the new method achieved state of the art on the EYEDIAP, MPIIFaceGaze, Gaze360 and RT-GENE datasets and achieved a performance increase of 0.02° to 0.30° compared to the other state of the art model. In addition, we show the generalization performance of the proposed model through a cross-dataset evaluation.

Axial spondyloarthritis (axSpA) is a chronic inflammatory disease of the sacroiliac joints. In this study, we develop a method for detecting bone marrow edema by magnetic resonance (MR) imaging of the sacroiliac joints and a deep-learning network. A total of 815 MR images of the sacroiliac joints were obtained from 60 patients diagnosed with axSpA and 19 healthy subjects. Gadolinium-enhanced fat-suppressed T1-weighted oblique coronal images were used for deep learning. Active sacroiliitis was defined as bone marrow edema, and the following processes were performed: setting the region of interest (ROI) and normalizing it to a size suitable for input to a deep-learning network, determining bone marrow edema using a convolutional-neural-network-based deep-learning network for individual MR images, and determining sacroiliac arthritis in subject examinations based on the classification results of individual MR images. About 70% of the patients and normal subjects were randomly selected for the training dataset, and the remaining 30% formed the test dataset. This process was repeated five times to calculate the average classification rate of the five-fold sets. The gradient-weighted class activation mapping method was used to validate the classification results. In the performance analysis of the ResNet18-based classification network for individual MR images, use of the ROI showed excellent detection performance of bone marrow edema with 93.55 ± 2.19% accuracy, 92.87 ± 1.27% recall, and 94.69 ± 3.03% precision. The overall performance was additionally improved using a median filter to reflect the context information. Finally, active sacroiliitis was diagnosed in individual subjects with 96.06 ± 2.83% accuracy, 100% recall, and 94.84 ± 3.73% precision. This is a pilot study to diagnose bone marrow edema by deep learning based on MR images, and the results suggest that MR analysis using deep learning can be a useful complementary means for clinicians to diagnose bone marrow edema.

Kinematic analysis of the hyoid bone in a videofluorosopic swallowing study (VFSS) is important for assessing dysphagia. However, calibrating the hyoid bone movement is time-consuming, and its reliability shows wide variation. Computer-assisted analysis has been studied to improve the efficiency and accuracy of hyoid bone identification and tracking, but its performance is limited. In this study, we aimed to design a robust network that can track hyoid bone movement automatically without human intervention. Using 69,389 frames from 197 VFSS files as the data set, a deep learning model for detection and trajectory prediction was constructed and trained by the BiFPN-U-Net(T) network. The present model showed improved performance when compared with the previous models: an area under the curve (AUC) of 0.998 for pixelwise accuracy, an accuracy of object detection of 99.5%, and a Dice similarity of 90.9%. The bounding box detection performance for the hyoid bone and reference objects was superior to that of other models, with a mean average precision of 95.9%. The estimation of the distance of hyoid bone movement also showed higher accuracy. The deep learning model proposed in this study could be used to detect and track the hyoid bone more efficiently and accurately in VFSS analysis.