Object Detection: Real-Time Road Damage Detection and Geolocation Using YOLOv8 and GNSS Integration

The road network serves not only as a key indicator of a city’s development but also as a foundational element in urban planning. As such, proper management is essential to ensure its effective functioning and sustainability [1]. Despite their importance, roads often present hazardous conditions for both humans and animals, prompting extensive research aimed at reducing accidents and improving road safety [2, 3]. In response to these challenges, conventional methods of road damage inspection—traditionally conducted manually and visually—have gradually evolved into automated approaches driven by Artificial Intelligence (AI) [4].

One of the most significant developments in AI is object detection in images and videos, which is being increasingly utilized in applications such as road surface monitoring. These methods use machine learning algorithms that can interpret complex data inputs, such as images, text, or audio [5-7]. Machine Learning (ML) offers a practical approach to achieving AI’s broader objective of extracting meaningful patterns from data, resulting in high accuracy across various domains, including image classification, facial recognition, and even human pose estimation [8-10].

Among the many algorithms developed for object detection, You Only Look Once (YOLO) stands out for its speed and efficiency [11, 12]. As a single-stage detector, YOLO has undergone several updates, including YOLOv3, YOLOv4, YOLOv5, and, most recently, YOLOv8, each offering significant performance improvements [13-16]. Earlier models, such as YOLOv5 and YOLOv7, achieved real-time detection with respectable accuracy; however, they often faced limitations in complex scenes involving small or overlapping objects—typical in road environments with varied lighting and surface conditions. Moreover, their modular architectures required additional customization for tasks like instance segmentation or object tracking. Ultralytics YOLOv8, released on January 10, 2023, addresses these limitations by offering an integrated, end-to-end architecture comprising a Backbone, Neck, and Head, optimized for object detection, classification, and instance segmentation [17, 18]. Compared to YOLOv5, YOLOv8 delivers improved precision, faster inference times, and native support for TensorRT and ONNX deployment, making it highly suitable for mobile and edge computing [19, 20]. Compared to two-stage detectors like Faster R-CNN, YOLOv8 offers significantly lower latency while maintaining competitive accuracy, an essential advantage for real-time road monitoring applications. This justifies the selection of YOLOv8 in this study, aiming to achieve a balance between speed, accuracy, and deployment flexibility.

However, while YOLO-based models are effective at identifying road damage types such as cracks, potholes, and patches, they do not inherently provide geographic information about the detected objects. In geospatial applications, position is typically represented using geographic coordinates, latitude and longitude, and in some cases, altitude. Without spatial context, the practical utility of detected road damage remains limited, particularly for tasks involving maintenance planning, navigation, or public reporting. Most existing studies on road damage detection using YOLO (e.g., YOLOv5 or YOLOv7) focus solely on detection accuracy, without addressing how to georeference the detected damage effectively [21]. This gap hinders the real-world usability of such systems, particularly in mobile and field-based deployments.

To address this gap, this study integrates YOLOv8 with Global Navigation Satellite System (GNSS) data to automatically tag detected damage with real-world geographic coordinates. GNSS provides real-time, global positioning capabilities that are unaffected by weather conditions, making it an ideal choice for outdoor data collection [22, 23]. Integrating GNSS with image and video capture enables seamless synchronization between visual data and spatial information. Android smartphones provide an ideal platform for this integration, offering support for both USB and wireless communication with external devices, such as cameras and GNSS receivers [24, 25]. Within this system, YOLOv8 is employed to detect road surface damage, while Optical Character Recognition (OCR) is used to extract coordinate information displayed on the screen. OCR tools, known for their high accuracy in recognizing Latin characters, facilitate this extraction process [26, 27].

In this context, the primary objective of the study is to develop an integrated system capable of detecting road damage while simultaneously capturing its spatial location in the form of geographic coordinates. The proposed solution simulates real-time object detection enriched with location tagging and further validates the accuracy of these coordinates through field-based GNSS measurements. This approach enhances the practicality and reliability of automated road damage detection systems, making it easier to identify, locate, and address road defects in real-world environments.

The design of the image recording device is developed by integrating several modular components that work together to enable the synchronized acquisition of image and positional data. The primary components in this system include a camera, a GNSS receiver, and a smartphone or tablet, which serves as the central processing and control unit. Each element plays a specific role in the data collection process: the camera captures real-time images or video of the road surface, the GNSS receiver provides accurate geographic coordinates, and the smartphone functions as both the controller and storage device. The connection between the camera and the smartphone is established via a USB-C interface, ensuring high-speed and stable data transfer for video streaming and control commands. Meanwhile, the GNSS receiver communicates wirelessly with the smartphone via Bluetooth, allowing for the flexible placement of the receiver module without the constraints of physical wiring. This modular and wireless design improves ease of use in field conditions, reduces clutter, and enhances mobility during data collection. The system is designed to be lightweight, portable, and adaptable, making it suitable for mobile surveying applications such as road condition monitoring, asset mapping, and geospatial data acquisition in dynamic environments.

GNSS circuit in Figure 2 is built utilizing the Beitian BN-220ZF GPS module, the ESP-32 IoT module, and the TP4056 Protect Charger. The circuit is created using a GNSS Module and an IoT-Nodemcu ESP-32 microcontroller. This Receiver GPS/GNSS device can stream NMEA positioning data wirelessly via Bluetooth. Smartphones receive a precision position streaming data acquisition application. This low-power, portable GNSS receiver design is optimized for field deployment where wired connections are impractical. By leveraging the wireless communication capabilities of the ESP-32 and the compact form factor of the BN-220ZF module, the system offers a flexible and efficient solution for integrating accurate geospatial data into mobile survey workflows. It is instrumental in road condition monitoring, asset mapping, and other field-based geoinformatics applications requiring centimeter-level positional accuracy.

tu_pian_2.png

Figure 2. GNSS circuit

The sturdy aluminum bracket serves as the central mount for various data acquisition devices installed on the car hood. Designed to be horizontal and symmetrical, this bracket supports the primary equipment. System stability is maintained by two strong suction cups, ensuring all devices remain secure even when the vehicle is in motion. The bracket’s design enables flexible and precise installation, making it ideal for road survey applications and automated field data collection.

tu_pian_3.jpg

Figure 3. Camera and GNSS receiver configuration on a bracket for road data acquisition

Figure 3 shows a camera mounted on one side of the support bar, as indicated by the left arrow, which records the journey or visually documents road conditions during data acquisition. On the opposite side, shown by the right arrow, a GNSS receiver is installed to obtain precise positioning data. Both devices are mounted on a horizontal bracket designed to provide stability and ensure optimal performance under dynamic conditions during field surveys.

A road data acquisition system is installed on a vehicle for visual surveying and position-based mapping. The system integrates several key components, including a USB camera, a GNSS receiver, a support bracket, and a smartphone, which serves as the monitoring device. The camera and receiver are mounted on the front hood of the vehicle using a horizontal metal bracket secured with suction cups to ensure device stability during vehicle movement. Inside the car, a smartphone is mounted on the dashboard, serving as a display unit to monitor the live video feed from the external camera in real-time.

tu_pian_4.jpg

Figure 4. In-vehicle monitoring of external camera feed for road surveying

The image in Figure 4 illustrates that the external camera is connected to the smartphone via a USB interface and controlled using a dedicated application. This application displays the live video stream from the camera, complete with a simple user interface, such as a “Record” button to start or stop recording. The GNSS receiver installed on the opposite side captures precise positional coordinates, allowing each video frame to be associated with accurate location data. This integrated setup enables efficient and synchronized road surveys, providing both visual and spatial documentation that can be used for various analyses, such as detecting road damage, mapping assets, or monitoring longitudinal environmental conditions.

The survey system produces data in the form of road surface images enriched with spatial information such as geographic coordinates (latitude and longitude), altitude, speed, and timestamp. This integration of visual and spatial data enables each captured image to be precisely georeferenced, allowing users to identify not only the type and severity of road surface conditions but also the exact location of the observed damage on a map. The inclusion of timestamp data ensures temporal tracking, which is essential for monitoring degradation over time or for comparing data across different survey periods. Meanwhile, vehicle speed is a critical factor affecting the reliability of recorded coordinates, as higher speeds can increase the margin of error in location measurement due to temporal displacement and GNSS lag. Altogether, this combination of attributes allows the dataset to support comprehensive spatial analysis, infrastructure planning, and maintenance decision-making, making it highly valuable for transportation authorities, urban planners, and geospatial analysts.

tu_pian_5.jpg

Figure 5. Detection image with additional coordinates

The image captured in Figure 5 shows the output of an automated road damage detection system that utilizes computer vision and geospatial tagging to identify and localize road surface defects. The image shows a pothole (“lubang”) detected with a confidence score of 0.87, enclosed in a bounding box and visually highlighted for straightforward interpretation. The system overlays metadata directly onto the video frame, including latitude, longitude, altitude, accuracy, and timestamp, enabling precise geolocation of the identified road damage.

Such a system combines visual analysis from a front-facing camera with GNSS data to support real-time condition monitoring of roads. The information displayed—such as coordinates and detection confidence—can be stored and later used for road maintenance planning, infrastructure audits, or integration into GIS-based asset management systems. This approach improves efficiency and consistency in road assessments, especially for large-scale urban or rural monitoring programs.

Furthermore, a position test was carried out. This test aims to obtain quantitative data regarding the position of the road damage class. Through this measurement, the information on the position or location of road damage, as shown by the model’s results in the application system, can be as precise or accurate. The use of the position or location generated by this application aims to facilitate identifying the damage site. Subsequently, it will make it easier for surveyors or those on duty to locate it again in the field.

The image shows a field verification activity conducted to directly identify road surface damage while simultaneously recording positional coordinates using a GNSS device with the NRTK (Network Real-time Kinematic) method. This method allows for centimeter-level positional accuracy, making it highly suitable for validating damage detection results from the automated system. Measurements are taken precisely at the damage point (in this case, a pothole), with the operator ensuring the device is held perpendicular to the surface and that the RTK correction signal remains stable. This process is essential to ensure the spatial reliability of the detection data. It serves as a reference for evaluating the performance of the image and AI-based mapping system used in the previous road survey.

The test compared the model-generated coordinates with the field verification results, revealing positional discrepancies between the automated detection system and the GNSS NRTK ground measurements. These differences were calculated as the deviation in latitude (ΔE) and longitude (ΔN), and the total positional error was computed in meters using the RMSE method. The resulting error values provide insight into the spatial accuracy of the automated detection system and serve as a critical reference for evaluating model performance and identifying potential improvements, whether in the detection algorithm or the integration of spatial data.

Table 1. Comparison of model and verified coordinates with positional error metrics

Table 1 presents the comparison results between the model-generated coordinates and the verified field measurements used to evaluate the positional accuracy of road damage detection. The ΔE (Delta Easting) and ΔN (Delta Northing) values represent the differences between the model and ground-truth coordinates, which were used to compute the total positional error using Euclidean distance. Figure 6 illustrates the ground verification process conducted using a GNSS RTK receiver, where the detected road damage points were precisely measured in the field to obtain high-accuracy reference coordinates. This procedure ensured that each detected defect was spatially validated against the model-generated coordinates, thereby supporting the positional accuracy assessment presented in Table 1. Across all observations, the RMSE was calculated at 5.523 m, with error values ranging from 0.670 m to 9.138 m. These errors reflect the overall spatial deviation and are influenced by factors such as the accuracy limitations of the GNSS module used (DGPS-grade), the relative distance between the camera and damage point, and the vehicle’s speed during data acquisition.

tu_pian_6.png

Figure 6. Implementation of verification of damage positions in the field

Further analysis revealed that vehicle speed has a measurable impact on geolocation accuracy. As detailed in Table 1 and visualized in Figure 7, higher speeds consistently resulted in greater positional discrepancies. Regression analysis confirmed a strong linear correlation between speed and positional error (R² = 0.7168), indicating that motion dynamics during capture can significantly degrade spatial accuracy. This result suggests that operating at lower speeds (≤ 30 km/h) can improve accuracy in mobile survey settings.

tu_pian_7.png

Figure 7. Plot of vehicle speed vs. RMSE

While the achieved RMSE does not satisfy sub-meter georeferencing standards, it is sufficient for practical road maintenance tasks where errors within a few meters remain visually traceable. Furthermore, a comparative evaluation with similar mobile mapping systems [38, 39] indicates that the system's performance aligns with industry-accepted accuracy levels for field-level planning and damage documentation.

These results confirm that controlling vehicle speed can enhance spatial accuracy in mobile surveys, and we recommend limiting speed to below 30 km/h for improved precision. The YOLOv8 model used for damage detection achieved reliable performance with mAP@0.5 of 76.4% and mAP@0.5:0.95 of 76.4% at an IoU threshold of 0.76, ensuring the reliability of visual detection outcomes.

Additionally, the study initially observed that the relative position of road damage within the image frame—specifically, the proximity of the damage to the camera—may influence detection clarity and, consequently, spatial accuracy. However, due to limitations in data collection, particularly the absence of direct measurements of object distance during field surveys, this relationship could not be quantitatively assessed. As such, we treat this influence as hypothetical in the present study. Prior literature in photogrammetry and geolocation has demonstrated that increased object distance can reduce image resolution and introduce geometric distortions, which may affect both detection performance and coordinate estimation. Notably, Dai et al. [40] highlight that spatial errors in photogrammetric measurements increase proportionally with object distance relative to the camera baseline, thereby degrading geolocation accuracy. In this context, we retain the discussion to contextualize possible sources of error observed in our spatial accuracy results. Figure 8 illustrates this concept by comparing two image captures of the same road defect taken from different distances. In subgraph (A) of Figure 8, the damage is recorded from a greater distance, while in subgraph (B), it is captured more closely. Although the damage and location are constant, the difference in visual appearance across frames likely influenced detection outcomes and could contribute to coordinate deviation.

tu_pian_8.jpg

tu_pian_9.jpg

Figure 8. Differences in the distance of road damage relative to the camera: (A) damage captured from a far distance, (B) damage captured from a closer distance within the image frame

The vehicle speed and distance between the camera and the actual damage location are critical factors that can affect the positional accuracy of data acquisition. This factor helps explain the spatial deviations observed, where differences between model-generated and verified coordinates reached several meters. Therefore, the relative position of the damage within the image frame should be considered a potential source of error in image-based geospatial data acquisition systems.