지능형 교통 체계를 위한 딥러닝 기반 객체 검출 활용 방법론 (Deep Learning based Object Detection Method and its Application for Intelligent Transport Systems)

서지정보

· 발행기관 : 제어·로봇·시스템학회
· 수록지 정보 : 제어.로봇.시스템학회 논문지 / 27권 / 12호 / 1016 ~ 1022페이지
· 저자명 : 김준용, 김성흠

초록

Most machine vision solutions for intelligent transport systems begin with the extraction of hand-crafted visual features from traffic scenes. In this study, the category-specific information for detecting specific visual patterns was investigated by replacing machine vision solutions with a standard convolutional neural network (CNN)-based object detector, and three practical applications of the system were developed. First, our system learned important categories, such as pedestrian and vehicles, for traffic monitoring and management. In addition, while collecting related databases, we efficiently performed data augmentation and improved the recognition accuracy of the system for several user-defined events. Further, the displacement of the detected positions between consecutive frames was converted into the real-world distance to compute the physical velocity of a vehicle. Second, we developed a vision-based system for a real-time lane-level traffic congestion measurement. After tracking the detected vehicles, the estimated velocities of vehicles for each lane were averaged. Subsequently, traffic congestion was determined based on the number of detected vehicles and averaged velocity. Third, we presented a context-aware method for background maintenance. To handle dynamic background objects, we utilized a 2D object detector to identify the category-specific background patterns. The key observation was that the detected regions do not belong to a true background. Hence, we developed a new confidence map to update the static background model and exclude pre-learnt background objects for conventional background subtraction methods. In the study, more than eight user-defined events were suggested by the combination of traditional machine vision techniques and deep learning-based object detectors with a substantial number of training images. In addition, our key ideas were validated using various datasets, such as five different scenes for lane-level traffic congestion and two CCTV image sequences for object-aware background subtraction and unseen object detection in the challenging traffic congestion. Lastly, the suggested applications of this system for intelligent transport systems were successfully demonstrated.

영어초록

Most machine vision solutions for intelligent transport systems begin with the extraction of hand-crafted visual features from traffic scenes. In this study, the category-specific information for detecting specific visual patterns was investigated by replacing machine vision solutions with a standard convolutional neural network (CNN)-based object detector, and three practical applications of the system were developed. First, our system learned important categories, such as pedestrian and vehicles, for traffic monitoring and management. In addition, while collecting related databases, we efficiently performed data augmentation and improved the recognition accuracy of the system for several user-defined events. Further, the displacement of the detected positions between consecutive frames was converted into the real-world distance to compute the physical velocity of a vehicle. Second, we developed a vision-based system for a real-time lane-level traffic congestion measurement. After tracking the detected vehicles, the estimated velocities of vehicles for each lane were averaged. Subsequently, traffic congestion was determined based on the number of detected vehicles and averaged velocity. Third, we presented a context-aware method for background maintenance. To handle dynamic background objects, we utilized a 2D object detector to identify the category-specific background patterns. The key observation was that the detected regions do not belong to a true background. Hence, we developed a new confidence map to update the static background model and exclude pre-learnt background objects for conventional background subtraction methods. In the study, more than eight user-defined events were suggested by the combination of traditional machine vision techniques and deep learning-based object detectors with a substantial number of training images. In addition, our key ideas were validated using various datasets, such as five different scenes for lane-level traffic congestion and two CCTV image sequences for object-aware background subtraction and unseen object detection in the challenging traffic congestion. Lastly, the suggested applications of this system for intelligent transport systems were successfully demonstrated.

참고자료

· 없음