Traffic Sign Mapping in Cambodia with Deep Learning and GNSS
Sophal Ratitya 1* and Mitsuharu Tokunaga2

1Graduate Student, Department of Civil and Environmental Engineering,
Kanazawa Institute of Technology, Japan
2Professor, Department of Civil and Environmental Engineering,
Kanazawa Institute of Technology, Japan
*titan.titya[at]gmail.com

Abstract

Traffic signs are necessary road features that help ensure safety. However, mapping these objects can be a time-consuming and tedious task. This study explores the use of deep learning and computer vision to improve the efficiency of traffic sign mapping in Cambodia. A traffic sign dataset was created from selected frames captured by a GoPro 12 action camera mounted on top of a car. These images were then labelled using image annotation software. Next, the traffic sign detection and classification models were trained using the pretrained YOLOv8 model in the Ultralytics framework. The Ultralytics tracking was used to maintain a unique ID of each sign in the video, allowing us to capture two frames as the object crosses a designated line. To determine the frame coordinate, we synchronize the video starting time with the Global Navigation Satellite System (GNSS) time. The next step is to calculate the detected sign in real-world coordinates. This involves finding the camera to GNSS difference based on camera height and field of view. The Haversine formula is used to find the distance between the camera in the pair frame. The objects^ pixel width in both frames is obtained from the detection model. The camera-to-object distance is calculated using the distance between the camera and the object^s pixel width, while the angle of the object to the camera^s line of sight is also measured. Finally, coordinates are calculated based on distance and angle. The experiments indicate that the approach accurately detects and identifies the coordinates of traffic signs with a mean absolute error of less than 5 meters. Consequently, the mapping process becomes easier and time-effective. Furthermore, an inventory can be easily built and updated frequently, facilitating efficient road asset management. While the system^s accuracy is mainly on the GNSS signal, it can be further improved using a high-precision GNSS integrated with other sensors or a real-time positioning (RTK) application. This research contributes to remote sensing by reducing the time needed for object mapping and by effectively integrating deep learning, computer vision, and GNSS for enhanced data acquisition.

Keywords: Computer Vision, Deep Learning, GNSS, Traffic Sign Mapping, YOLOv8

Topic: Topic C: Emerging Technologies in Remote Sensing