Hiroyuki Kumazawa

Drones having high resolution cameras and sensors have become more available and cheaper. We use drone-captured images taken directly above the plantation to research how well the types and areas of fruit trees, as well as other features, can be recognized. Wide-area images are synthesized from numerous locally captured images taken by the drone and are then divided into blocks (image blocks) with a certain amount of overlap to increase the number of blocks available for training. In this paper, semantic segmentation is applied to these blocks, and their classification performance is evaluated. In these evaluations, it is clarified that the overlaps in the blocks make it difficult to properly separate the training data for training the semantic segmentation network from the test data for performance evaluation. To address these issues, data augmentation is applied to the test data, and the evaluation results are presented.

In ad-hoc network such as inter-vehicle communication (IVC) system, safety applications that vehicles broadcast the information such as car velocity, position and so on periodically are considered. In these applications, if there are many vehicles broadcast data in a communication area, congestion incurs a problem decreasing communication reliability. We propose autonomous distributed transmit power control method to keep high communication reliability. In this method, each vehicle controls its transmit power using feed back control. Furthermore, we design a communication protocol to realize the proposed method, and we evaluate the effectiveness of proposed method using computer simulation.

We propose a general prediction method based on the efficient computation and online update of the Singular Value Decomposition (SVD) of historical data. The SVD is fundamental to many data modeling algorithms, but the traditional methods for computing it require large computational costs. By adopting a fast sequential SVD updating scheme, the tasks of prediction, imputation of missing values, and model updating can be performed very quickly. In this paper, an application of our method to route travel time prediction is described. Using real travel time data from short sections (links) on expressway, we evaluated prediction performance of travel time on longer section (route) including the links. Experimental comparisons with several statistical machine learning methods suggest that our linear prediction method can achieve similar prediction performance (prediction error) to other nonlinear methods at less computaional cost.