In recent years, the number of fatalities in traffic accidents involving motorcyclists has remained almost unchanged, with single-vehicle accidents accounting for 37.2% of all accidents by accident type in the past five years. In the development of overturn prevention devices for motorcycles, problems remain in post-mounting of the device as well as its downsizing. On the other hand, an existing study using deep learning has proposed a method for detecting dangerous objects on the road surface leading motorcycles to overturn, though this method still needs verification under different conditions. In this study, we apply a method for detecting dangerous objects on the road surface from video images using YOLO to two types of 360-degree cameras and verify that this method is versatile under different conditions.

View full abstract

This paper presents a custom-built IoT camera system designed for recognizing wild animal approaches, where data transmission and power consumption are critical concerns in resource-constrained outdoor settings. The proposed method involves the spectral analysis on both infrared and environmental sound data before uploading images and videos to the remote server. Experiments, including battery endurance tests and wildlife monitoring, were conducted to validate the system. These results showed that the system minimized false positives caused by environmental factors such as wind or vegetation movement. Importantly, adding frequency features from audio waveforms that capture sounds including wind noise and footsteps led to an improvement in detection accuracy, which increased the AUC from 0.894 to 0.990 in Random Forest (RF) and from 0.900 with infrared sensor data alone to 0.987 in Logistic Regression (LR). These findings contribute to applications in wildlife conservation, agricultural protection, and ecosystem monitoring.

View full abstract

In Japan, the shortage of human resources due to the declining birthrate and aging population is becoming a social problem. Particularly in the security industry, the irregular working hours and associated risks are making it increasingly challenging to secure workers. This has led to a rise in use of security systems that utilize security cameras and drones. However, in factories and other buildings with a lot of equipment and intricate structures, there is the problem of blind spots caused by occlusion. This situation necessitates the use of automated drone patrols, and a problem arises when self-position estimation fails in areas where acquiring feature points is difficult, such as corridors. To solve these problems, in a previous study, we devised a technique for position estimation using a method that can calculate similarity based on changes in the distribution of color information across the entire image. In this study, we propose a method that can cope with environmental changes caused by object movement while combining feature point-based methods.

View full abstract

The turning movement count is investigated to understand the traffic conditions at intersections and identify bottleneck locations. In recent years, methods utilizing probe data and AI-based analysis of video images have been developed to streamline the survey process. Existing methods can count vehicles as they pass but struggle to classify vehicle types. Therefore, the objective of this study is to develop a method for counting turning movement count by vehicle type using deep learning. In this method, YOLOv8 is used to detect cars, buses, and trucks in video images, and BoT-SORT is used for tracking. When a vehicle being tracked crosses the cross-sectional lines and auxiliary lines at the intersection captured in the video images, it is counted by class. In this case, the entry direction of vehicles that cannot be determined upon entering the intersection is estimated based on accurately counted vehicles. Additionally, the entry direction is inferred from a series of vector information within the detection bounding boxes. The results of the verification experiment showed that the proposed method can count the directional traffic volume with an accuracy of over 95.0% and classify the three vehicle classes—car, bus, and truck—with an accuracy of over 90.0%.

View full abstract

Labor shortages in the construction industry have become a serious issue in developed countries, particularly in Japan, where workforce aging and declining recruitment of young workers are significant challenges. In this context, ensuring worker safety has become increasingly critical. While occupational accidents in Japan's construction industry have decreased annually due to proper safety measures, the construction industry still has the highest number of fatalities among all industries. Falls from height and falls on the same level are the leading causes of injuries and fatalities. Therefore, detecting near-miss incidents (such as tripping and slipping) that precede falls, along with physical fatigue, could help prevent occupational accidents. This study investigated the feasibility of detecting near-miss incidents and estimating fatigue levels using wearable sensors suitable for continuous monitoring at construction sites. We conducted validation experiments simulating near-miss actions and fatigue conditions. Results showed that applying a Convolutional Neural Network (CNN) to data collected from an iPhone® placed in workers' trouser pockets achieved an F1-score of 0.95 in detecting near-miss actions. Additionally, by comparing body sway magnitudes before and after fatigue, we confirmed the potential for estimating physical fatigue.

View full abstract