Hiroyuki Kumazawa

Drones having high resolution cameras and sensors have become more available and cheaper. This makes it possible to use them to take images at close range from directly above, which would normally be very hard to take. We used the images taken by a drone from directly above a plantation to research on how well the types and area the fruit trees and other features occupy can be recognized. We reported the application of the deep learning, specifically semantic segmentation, for the monitor of the land usage, which has high accuracy and outstanding results in recognizing objects in the images. In this paper, classification accuracies are evaluated for the various parameters of image blocks, which are extracted from wide area images and used as training data set for the semantic segmentation.

Drones having high resolution cameras and sensors have become more available and cheaper. This makes it possible to use them to take images at close range from directly above, which would normally be very hard to take. We used the images taken by a drone from directly above a plantation to research on how well the types and area the fruit trees and other features occupy can be recognized. In this paper, we apply the deep learning, specifically semantic segmentation, for the monitor of the land usage, which is pixel-by-pixel analysis of an image and decides what class each pixel belongs to. That means the fine classification is possible. In this paper, we show how to capture images and what can be realized using such images. Then classification results of the types and area the fruit trees and other features are shown by using semantic segmentation with various image block sizes.

We have been considering methods to detect transportation modes using sensor data such as GPS, acceleration, and gyroscope data collected from smartphones. Transportation modes include walk, vehicle, bus, train, and bike. In the previous papers, we presented transportation mode detection from acceleration and gyroscope data using machine learning such as decision tree and random forest, and the method for correcting the classification errors by the post processing of the results of machine learning, which leads to the improvement of the accuracy of the detection. In this paper, we consider the application of LSTM (Long Short Term Memory), which can deal with time series data, aiming at replacing the post processing. At this moment, the LSTM cannot attain the performance of the post processing due to many factors such as the setting of many parameters of LSTM, feature values, and so on, which will be the future works.

In this study, we propose a method for fingerspelling recognition using both color and depth images by using an RGB-D camera and a Multi Modal Convolutional Neural Network (MMCNN) for automatic fingerspelling translation. In the proposed method, two pre-processing steps are performed on the depth image: distance normalization and background removal more than 1 meter from the camera. Experimental results show that the recognition rate of the proposed method outperforms that of the method without preprocessing or with either of the two preprocessing methods.

Drones having high resolution cameras and sensors have become more available and cheaper. This makes it possible to use them to take images at close range from directly above, which would normally be very hard to take. This characteristics can be applied to many fields such as supervisory and control, management of assets in the wide area, and so on. Recently, different technologies are being used to improve management and increase output in the farming industry, one of these methods is the application of drones. We used the images taken by a drone from directly above a plantation to research on how well the types and area the fruit trees are planted can be classified. In this paper, we applied the deep learning with CNN (Convolutional Neural Network) for the classification, which has high accuracy and outstanding results in recognizing objects in the images.

We have been considering methods to detect transportation modes using sensor data such as GPS, acceleration, and gyroscope data collected from smartphones. Our methods use machine learning such as decision tree and k-nearest neighbor algorithms to design classifiers, which classify feature values calculated from the collected data into one of the transportation modes. However, the classification errors are inevitable for machine learning, which leads to increasing the detection error rate. Therefore, we proposed the modification method to reduce the detection errors by the post processing of machine learning. Our classification and modification methods use GPS as a key information, which restricts the availability of the methods within the GPS receivable areas. To overcome this situation, we consider the methods, which do not require GPS information.

Face recognition system is considered using machine learning and face recognition API (Application Programming Interface) provided by the web services. The system uses an omnidirectional camera located in the fixed position to take pictures of the laboratory room, detects faces of people sitting in the room, and checks them by comparing with the pre-registered faces in the database. This paper describes the implementation of the prototype system. The elementary results of the operation of the prototype system revealed various problems, which are caused by image qualities, the required time for data transfer and processing, and so on. Among them, the reduction of the processing time is considered.

Recent deployment of smartphones and car navigations makes it possible for mobile entities such as people and vehicles to have sensors and communication capabilities, and to send their sensor data by themselves during moving. To make these collected data valuable, how to analyze these sensor data is an important issue. We have been considering methods to detect transportation modes using sensor data such as GPS, acceleration, and gyroscope data collected from smartphones. Our methods use machine learning such as decision tree and k-nearest neighbour algorithms to design classifiers, which classify feature values calculated from the collected data into one of the transportation modes. However, the classification errors are inevitable for machine learning, which leads to increasing the detection error rate. Considering the moving processes of travellers, we assume that the transportation mode doesn’t change frequently in a short period of time. Under this assumption, the method for correcting the classification errors is proposed and its improvement of the accuracy of the detection is shown.

Machine learning technologies have been progressing and are becoming commodities, recently. In addition to many frameworks for the implementation of the technologies, the emergence of the web services for using the technologies is making their utilization easier and easier. Under these circumstances, face recognition system is considered using machine learning and face recognition API (Application Programming Interface) provided by the web services. The system uses an omnidirectional camera located in the fixed position to take pictures of the laboratory room, detects faces of people sitting in the room, and checks them by comparing with the pre-registered faces in the database. This paper describes the implementation of the prototype system and the elementary results of the operation of the prototype system. The implementation of the system revealed various problems which are caused by image qualities, the required time for data transfer and processing, and so on. Some countermeasures against the problems are also described in the paper.

The decision trees for machine learning are designed to detect the transportation modes which travelers use, that is, trains, cars, buses, bikes, and walking, and their precisions are evaluated. Then the method for correcting the detection errors by the weighted majority vote is proposed and its precisions are shown.

We have been considering methods to detect transportation modes using sensor data such as GPS, acceleration, and gyroscope data collected from smartphones. Our methods use machine learning such as decision tree and k-nearest neighbour algorithms to design classifiers, which classify feature values calculated from the collected data into one of the transportation modes. When the machine learning classifiers classify the feature values into the specific transportation mode, the classification errors are inevitable. Considering the moving processes of travellers, we assume that the transportation mode doesn’t change frequently in a short period of time. Under this assumption, the method for correcting the classification errors is proposed and its improvement of the accuracy of the detection is shown. Furthermore, we observe the occurrence tendency of the detection errors and what kind of changes occur before and after the correction processing. As the result, the characteristics of the detection errors and how the correction processing works in the detection are shown.

We have been considering methods to detect transportation modes using sensor data such as GPS, acceleration, and gyroscope data collected from smartphones. Our methods use machine learning algorithms to design classifiers, which classify feature values calculated from the collected data into one of the transportation modes. In the real applications, the designed classifiers are implemented on the terminal side or the central server side according to the services. Then the feature values are fed into the classifiers. When the implemented classifiers classify the feature values into the specific transportation mode, the classification errors are inevitable. Considering the moving processes of travellers, we assume that the transportation mode doesn’t change frequently in a short period of time. Under this assumption, the method for correcting the classification errors is proposed and its improvement of the accuracy is shown. Furthermore, we observe the occurrence tendency of the detection error and show what kind of changes occur before and after the correction processing. As the result, how the correction processing works in the detection is shown.

Probe information system for vehicles using smartphone is implemented. In this system, CAN data and smartphone sensor including its camera information are transmitted to server through communication capabilities of the smartphone. These probe data are stored in the database and delivered to web terminals through Web API and push type delivery mechanism.

The decision trees for machine learning are designed to detect the transportation modes which travelers use, that is, trains, cars, buses, bikes, and walking, and their precisions are evaluated. Then the method for correcting the detection errors by the designed decision tree is proposed and its precisions are shown.

The processing of sensor data such as GPS, acceleration, and gyroscope data collected from smartphones are shown to detect transportation modes. Transportation modes are the means which travellers use, that is, trains, cars, buses, bikes, walking, and so on. First, the compensation method of the orientation of a smartphone is described, because the orientation of a smartphone affects sensor data, especially acceleration data. The compensation method transforms the coordinate system of a smartphone to the reference coordinate system whose coordinate axes are vertical and horizontal to the ground. As the result, sensor data along the vertical and horizontal directions are always acquired. Then, machine learning such as decision tree and k-nearest neighbour algorithms is applied to the feature values calculated from the collected data and their evaluation results of detecting the transportation modes are shown.

The processing of sensor data such as GPS, acceleration, and gyroscope data collected from smartphones are shown to detect the transportation modes. The transportation modes are the means which travelers use, that is, trains, cars, buses, bikes, walking, and so on. First, the compensation method of the orientation of the smartphone is described, because the orientation of the smartphone affects sensor data, especially acceleration data. The compensation method transforms the coordinate system of the smartphone to the fixed reference coordinate system whose coordinate axes are vertical and horizontal to the ground. As the result, sensor data along the fixed directions are always acquired. Then, machine learning technologies such as decision tree and k-nearest neighbor methods are applied to the compensated data and their evaluation results of detecting the transportation modes are shown.