http://chineseinput.net/에서 pinyin(병음)방식으로 중국어를 변환할 수 있습니다.
변환된 중국어를 복사하여 사용하시면 됩니다.
Prediction of Dangerous Pedestrians using Depth and Stance Estimation
Stephanie Nix,Kana Koishi,Hirokazu Madokoro,Takashi K. Saito,Kazuhito Sato 제어로봇시스템학회 2022 제어로봇시스템학회 국제학술대회 논문집 Vol.2022 No.11
According to the Tokyo Fire Department, from 2015 to 2019, there were 211 people who were taken to the hospital due to accidents involving pedestrians using their smartphones while walking. Around 40% of these accidents were due to pedestrians running into bicycles, people, and other objects. Local ordinances have been passed to reduce the danger posed by pedestrians engaging in this unsafe behavior, but this has not significantly reduced the use of smartphones while walking. In this paper, we propose a metric for evaluating the level of danger posed by pedestrians captured on video. In our proposed method, we extract the skeletons of pedestrians using MediaPipe, then predict the danger level by estimating the pedestrian stance and depth within the image. Then, we evaluate the accuracy of our model by calculating the accuracy of the depth and stance estimators on a video taken by a car-mounted camera. We estimated danger levels on a local road and obtained an accuracy of 0.459.
Hirokazu Madokoro,Saki Nemoto,Stephanie Nix,Osamu Kiguchi,Atsushi Suetsugu,Takeshi Nagayoshi,Kazuhito Sato 제어로봇시스템학회 2022 제어로봇시스템학회 국제학술대회 논문집 Vol.2022 No.11
Air pollution causes various health problems and diseases. Long-term PM<SUB>2.5</SUB> monitoring and prediction of its occurrence and sources are necessary not only in global areas based on public monitoring stations but also in local areas using cost-effective sensor systems. For this study, we developed a sensor system to achieve simplified and high-frequency PM<SUB>2.5</SUB> measurements. We attempted to learn and to predict local PM<SUB>2.5</SUB> concentrations from observed data using long short-term memory (LSTM) as a dominant time-series feature learning network. For improving learning and prediction accuracy evaluated according to the root mean square error (RMSE), sensor calibration is performed using a higher sensor. Moreover, we strove to reduce RMSE by optimizing its five major parameters. Experimentally obtained results demonstrate that the prediction accuracy is improved gradually after calibration and parameter optimization. As an ablation experiment, five meteorological factors are imported externally to verify the factors which contribute to reducing RMSE. Results verify the strong effects of local pressure and temperature for training and relative humidity and temperature for testing as validation.
Readiness Estimation for a Take-Over Request in Automated Driving on an Expressway
Ryohei Suzuki,Hirokazu Madokoro,Stephanie Nix,Kazuki Saruta,Takashi K. Saito,Kazuhito Sato 제어로봇시스템학회 2022 제어로봇시스템학회 국제학술대회 논문집 Vol.2022 No.11
Automated driving is attracting attention as a solution to road traffic problems. At Level 3, a take-over request (TOR) is issued to transfer driving operations from the system to a driver when it is unable to continue. In such cases, the driver must be monitored to ensure a proper takeover of the driving operations. This study aims to measure drivers’ brain activity before and after the TOR by analyzing time-series signals of brain activity with machine learning algorithms. We developed driving scenarios with a TOR trigger on a rainy expressway at night. We used a portable functional near-infrared spectroscopy (fNIRS) device to measure cerebral blood oxygenation changes (ΔHbO) at the frontal pole. We used a long short-term memory (LSTM) network on this data for time-series learning and prediction after multivariate and multilayering modifications to improve accuracy. We conducted driving questionnaires beforehand and used two classification methods to categorize subjects into several groups with similar driving characteristics. Experimental results of a ΔHbO drop revealed that brain activity tended to decrease during automated driving. Moreover, success in obstacle avoidance and mean squared error (MSE) for each driver group demonstrated that the behavior toward an obstacle after the TOR trigger influenced changes in brain activity.
Yuto Shoji,Hirokazu Madokoro,Stephanie Nix,Kazuki Saruta,Takashi K. Saito,Kazuhito Sato 제어로봇시스템학회 2022 제어로봇시스템학회 국제학술대회 논문집 Vol.2022 No.11
The number of accidents involving pedestrians and bicyclists has been reported to be about 1.8 times higher on narrow roads than on arterial roads in Japan. We consider investigating the circumstances under which accidents occur on narrow roads to be an important research task. Statistics from the Tokyo Metropolitan Police Department indicate that the number of traffic accident fatalities in winter is relatively high. We used a Driving Simulator (DS) in order to safely perform sensing on roads that replicate a local city in a heavy snowfall region. Brain activity during driving was measured using a portable functional Near-Infrared Spectroscopy (fNIRS) device. We used a machine-learning algorithm for analyzing time-series datasets to demonstrate differences in brain activity across driving events. We classified subjects into four groups based on the results of questionnaires that assessed their driving characteristics. Experimentally obtained results demonstrated that Root Mean Squared Error (RMSE) changes that represent increased brain activity were greater in winter than in summer for each event. We infer that the winter events had a larger impact on the drivers.
Calibration and 3D Reconstruction of Images Obtained Using Spherical Panoramic Camera
Hirokazu Madokoro,Satoshi Yamamoto,Yo Nishimura,Stephanie Nix,Hanwool Woo,Kazuhito Sato 제어로봇시스템학회 2021 제어로봇시스템학회 국제학술대회 논문집 Vol.2021 No.10
This study was conducted to develop a 3D reconstruction procedure for application to crop monitoring. For 3D construction of a similar target object, we compared images obtained from two camera types: a compact digital camera (CDC) and a spherical panoramic camera (SPC). First, we calculate camera parameters from images that include a checkerboard. Subsequently, we correct the image distortion including that of the target object using the camera parameters. Finally, we estimate camera positions and three-dimensional (3D) reconstruction based on the structure from motion (SfM). Experimentally obtained results demonstrated that the 3D reconstruction of a target object was improved after calibration compared with that before calibration. Moreover, we conducted an application experiment using a tree in an outdoor environment as a trial of practical use at a farm.
Kodai Sato,Hirokazu Madokoro,Takeshi Nagayoshi,Shun Chiyonobu,Paolo Martizzi,Stephanie Nix,Hanwool Woo,Takashi K. Saito,Kazuhito Sato 제어로봇시스템학회 2021 제어로봇시스템학회 국제학술대회 논문집 Vol.2021 No.10
This study was conducted to classify outcrop images using semantic segmentation methods based on deep learning algorithms. Carbon capture and storage (CCS) is an epoch-making approach to reduce greenhouse gases in the atmosphere. This study specifically examines outcrops because geological layer measurements can lead to production of a highly accurate geological model for feasible CCS inspections. Using a digital monocular RGB camera, we obtained 13 outcrop images annotated with four classes along with strata. Subsequently, we compared segmentation accuracies with changing input image sizes of three types and semantic segmentation methods of four backbones: SegNet, U-Net, ResNet-18, and Xception-65. The ResNet-18 and Xception-65 backbones were implemented using DeepLabv3+. Experimentally obtained results demonstrated that data expansion with random sampling improved the accuracy. Regarding evaluation metrics, global accuracy and local accuracy are higher than the mean intersection over union (mIoU) for our outcrop image dataset with unequal numbers of pixels in the respective classes. These experimentally obtained results revealed that resizing for input images is unnecessary for our method.
Domain Adaptation for Agricultural Image Recognition and Segmentation Using Category Maps
Kota Takahashi,Hirokazu Madokoro,Satoshi Yamamoto,Yo Nishimura,Stephanie Nix,Hanwool Woo,Takashi K. Saito,Kazuhito Sato 제어로봇시스템학회 2021 제어로봇시스템학회 국제학술대회 논문집 Vol.2021 No.10
Recognition accuracy obtained using deep learning drops precipitously when the training data are insufficient. This paper presents a data-expansion method for training of the transfer learning source domain. Using expanding images generated from weights on a category map as source data, we compared accuracies obtained from five derivative models and our previously reported method. Moreover, we obtained the result of domain adaptation between actual images and synthetic images using weights obtained during transfer learning. Based on those results, we verify whether the amount of training data can be expanded quantitatively and qualitatively. Experiment results obtained from two open benchmark datasets and our original benchmark dataset demonstrated that our proposed method outperforms the previous method under a guarantee of sufficient accuracy for the synthetic images.