Road Pothole Detection Using Yolov8 Custom Segmentation

Road Pothole Detection Using Yolov8 Custom Segmentation

Neha, Murari B K

PG Student, St. Joseph Engineering College, Mangalore Assistant Professor, St. Joseph Engineering College, Mangalore

Abstract – We propose a real-time pothole identification and alert system powered by a tailored YOLOv8 instance segmentation approach. The system is designed to enhance road safety by accurately detecting potholes from various input sources, including static images, recorded videos, and live dashboard camera feeds. Trained on a custom dataset, the model identifies potholes and outlines their contours to determine precise location and size. The system logs both the quantity of identified potholes and their precise geographic coordinates through geolocation technology. An alert message containing the relevant details is then generated and forwarded to the concerned authorities for prompt action. The system operates with minimal human intervention, offering a fully automated and scalable solution for urban road monitoring. Additionally, the annotated outputs can assist in maintenance planning by highlighting frequently affected zones. The experimental findings prove the models reliability, indicating promising integration with intelligent transport systems to reduce hazards and streamline infrastructure upkeep.

Keywords: Pothole Detection, YOLOv8, Geolocation, Automated Road Monitoring

1. INTRODUCTION

   Potholes are a widespread issue that affects roads globally. They dont just make drives bumpy and uncomfortable – they can also damage vehicles and even lead to serious accidents. Right now, most cities find potholes in one of two ways: either by sending crews to inspect roads (which takes time and manpower) or waiting for drivers to report them (which means many get missed).

   Our solution changes this with smart technology that spots potholes automatically. Using a powerful AI system called YOLOv8, it doesnt just notice potholes exist – it can map their exact shape and depth, giving far more detail than older methods. The technology accepts visual input from multiple sources, including still images, pre-recorded videos, and live streams from vehicle-mounted or infrastructure cameras.

   When it finds a pothole, it instantly logs the GPS location, counts how many are in the area, and alerts maintenance teams with all the details. This cuts out slow manual reporting and gets repairs started faster. After rigorous testing with real road data, weve proven the system is both accurate and dependable, perfect for modern city planning. Plus, all the collected data helps cities spot problem areas and plan better road maintenance long-term.

2. LITERATURE REVIEW

   Recent progress in computer vision and artificial intelligence has dramatically improved automated pothole identification systems. A comprehensive study on YOLOv8-based detection

   methods [1] evaluated multiple existing techniques, examining YOLOv3 and YOLOv4 architectures for dimensional analysis, Raspberry Pi-powered IoT solutions for immediate detection, a combined CNN-SVM-YOLO framework achieving 82percent precision, and R-CNN models incorporating dilated convolutional layers. This research established YOLOv8s advantages in detection accuracy, processing speed, and performance consistency across variable road conditions.

   In related work, Khare and colleagues [2] performed extensive comparisons between YOLOv8 and its predecessors (v5 and v7) for identifying various road hazards including potholes, drainage covers, and access points. Their findings verified YOLOv8s enhanced precision, faster computation times, and greater reliability in practical scenarios. The research particularly stressed how strategic data enhancement and careful parameter optimization significantly boost system performance, positioning YOLOv8 as an effective choice for instant hazard recognition.

   An alternative approach was presented by Parasnis et al. [3] through their innovative RoadScan framework. This solution tackles the computational constraints of conventional deep learning models like YOLO and SSD by integrating VGG16 architecture with a Siamese neural network and triplet loss function. Their lightweight design attained exceptional results (96.12percent accuracy with just 3.89percent error rate) while minimizing processing requirements, making it particularly valuable for hardware-limited applications.

   Subsequent research [4] analyzed YOLOv8s detection capabilities when supplemented with different image optimization methods. Through evaluation of camera-based AI detection against instrument-based approaches (including accelerometer-GPS systems), the research determined that while sensor methods are budget-conscious, they exhibit inconsistent performance and device dependencies. The experiments confirmed that YOLOv8, when optimized with techniques such as exposure modification and rotational augmentation, delivers consistent and precise pothole identification regardless of environmental variables.

3. PROPOSED METHODOLOGY

   Our technical approach for building a live pothole monitoring solution follows five sequential phases: gathering road condition data, preprocessing visual inputs, training the detection algorithm, evaluating model effectiveness, and deploying the functional system.

   1. Data Collection

      • The research utilized an extensive collection of roadway images sourced from Kaggle, containing examples of both deteriorated pavement with potholes and undamaged road surfaces, photographed under diverse illumination and meteorological conditions. The images were manually annotated using segmentation masks to accurately outline pothole boundaries, allowing the model to learn both the presence and precise shape of potholes.

   2. Data Preprocessing

      All collected images were resized to a uniform resolution of 640×640 pixels to ensure compatibility with the YOLOv8 segmentation model. To improve the models ability to perform well on unseen data and reduce overfitting, we applied various image transformations including horizontal flips, brightness/contrast variations, rotations, and controlled noise injection.

   3. Model Training

      The YOLOv8 segmentation model, known for its real-time performance and accuracy, was selected for the detection task. The model was configured with a batch size of 16, a learning rate of 0.0001, and trained for 150 epochs. The collected data was partitioned with 80 percent allocated for model training and the remaining 20percent reserved for evaluation purposes. Training focused on joint optimization of three distinct objectives: accurate object classification, precise localization coordinates, and detailed instance segmentation masks.

   4. Model Evaluation

      • To assess the models effectiveness, several evaluation metrics were used:

      • Accuracy: Achieved a validation accuracy of 95percent.

      • Precision and Recall: Two critical metrics were analyzed: precision (indicating how many reported potholes were genuine) and recall (showing what percentage of existing potholes were successfully detected).

      • F1-Score: For a balanced measure of precision and recall.

      • Frames Per Second (FPS): Ensured the model operates in real time with a processing rate suitable for video streams.

   5. System Integration and Deployment

      After training, the model was integrated into a real-time detection system capable of processing static images, recorded videos, and live dashboard camera feeds. Upon detecting potholes, the system outlines their shapes, counts the number of potholes, and uses GPS metadata to locate ther positions. It then automatically generates a detailed alert report, including timestamp, coordinates, pothole count, and a segmented image. This report is sent instantly to road maintenance authorities via a communication platform (e.g., Telegram bot) for timely repair response.

      Fig. 1: System architecture

      In Figure 1 presents the complete system architecture combining a YOLOv8 object detection model with a Flask web application for pothole detection. The workflow starts with collecting and preprocessing training images through data augmentation techniques to improve model generalization. The images receive labeling before being utilized to teach the YOLOv8 model to identify pothole characteristics. After successful training, the model gets integrated into a Flaskbased web interface that allows users to submit images for analysis. When a user uploads a road image, the trained model processes it in real-time, identifies any potholes present, and returns the results through the web interface with visual bounding boxes marking detected potholes. This end-to-end system demonstrates practical implementation of computer vision for infrastructure monitoring applications.

4. IMPLEMENTATION

   • Pothole detection model is embedded into an operational system that automatically examines photos, pre-recorded footage, and streaming video from dashboard cameras in real-time.

   • GPS Data Extraction When a pothole is found, the system retrieves its location from the medias metadata or a connected GPS device.

   • Pothole Analysis It counts the potholes, estimates their sizes from the segmented areas, and categorizes severity. Image Capture A image is saved with the detected potholes highlighted for visual confirmation. The system automatically saves a still image with clearly marked pothole boundaries, providing visual evidence of each detection.

   • Upon detection, the system automatically compiles a detailed report containing the exact time, geographic coordinates, number and severity of potholes, along with visual evidence, then immediately transmits this alert via Telegram to maintenance crews while simultaneously logging all findings in a centralized database to identify recurring problem areas and optimize repair prioritization

5. RESULT AND DISCUSSION

   The model achieved an accuracy of 95percent, indicating that it correctly identified potholes and non-potholes in 95percent of the test cases. A precision of 0.92 means that when the model predicted a pothole, it was correct 92percent of the time demonstrating a low rate of false positives. The recall of 0.85 shows that the model was able to detect 85percent of all actual

   potholes, although it missed a few (false negatives). The F1- score of 0.88, which balances both precision and recall, confirms that the model performs reliably in both identifying potholes and minimizing incorrect detections. The findings demonstrate a stable and well-optimized model ready for practical implementation. These metrics demonstrate the models strong ability to generalize across diverse road conditions and varying image inputs. Overall, the systems consistent performance highlights its readiness for integration into intelligent transportation and automated maintenance workflows.

   TABLE I: YOLOv8 Model Performance Metrics

   Metric	Value (%)
   Accuracy	95.0
   Precision	0.92
   Recall	0.85
   F1 Score	0.82

   1. Training and validation

      Fig. 2: Training and validation

      In Fig.2 The YOLOv8 segmentation model, known for its real- time performance and accuracy, was selected for the detection task. The model was configured with a batch size of 16, a learning rate of 0.0001, and trained for 150 epochs. The collected data was partitioned with 80 percent allocated for model training and the remaining 20percent reserved for evaluation purposes. Training focused on joint optimization of three distinct objectives: accurate object classification, precise localization coordinates, and detailed instance segmentation masks.

   2. Model evaluation metrics

      Fig. 3: Precision recall and F1 score

      The Precision-Recall and F1 Score curves in Figure 3 illustrate the models accuracy in pothole detection. The blue curve represents how precision varies as recall increases, illustrating the typical trade-off between identifying all actual potholes (recall) and maintaining the correctness of those detections

      (precision). Initially, the model maintains high precision, but as recall increases, precision begins to drop, which is expected in most classification systems.

      The dashed purple curve shows the corresponding F1 Score, which combines both precision and recall into a single metric. The graph supports the conclusion that the model maintains consistent performance and is well-suited for real-world pothole detection applications. This balanced score highlights the models ability to minimize both false positives and false negatives. As evidenced by the graph the model exhibits stable detection capabilities, making it appropriate for deployment in actual road inspection scenarios

   3. Testing result

      Fig. 4: Image of the road with detected pothole

      The test output from the YOLOv8 model demonstrates successful multi-pothole detection within a single image frame, achieving confidence scores. Each identification is precisely bounded by pothole labeled boxes, verifying the models classification and localization competence across challenging terrain conditions. The consistently high confidence metrics and distinct, non-overlapping bounding boxes validate both prediction reliability and object differentiation capability.

      The dashed purple curve shows the corresponding F1 Score, which combines both precision and recall into a single metric. The graph supports the conclusion that the model maintains consistent performance and is well-suited for real-world pothole detection applications. This balanced score highlights the models ability to minimize both false positives and false negatives. As evidenced by the graph the model exhibits stable detection capabilities, making it appropriate for deployment in actual road inspection scenarios

6. CONCLUSION

   Our research team has successfully created an intelligent road monitoring solution that combines advanced camera systems with artificial intelligence to automatically detect various types of pavement damage like potholes, cracks, and surface wear. The specially designed detection software analyzes different kinds of visual data – including photos from smartphones, footage from dashboard cameras, and live feeds from cameras – quickly identifying problems with reliable accuracy. The system instantly records the exact location of every detected road issue using the vehicles built-in GPS, ensuring high-precision positioning. Immediately upon identifying road damage, the system produces complete digital documentation containing: the precise moment of detection, accurate GPS positioning data, a categorized assessment of seriousness (ranging to critical)and annotated photos highlighting the damage – all sent

   automatically through popular messaging platforms directly to the responsible maintenance departments.

   This innovative approach delivers three significant improvements to traditional road maintenance: firstly, it completely eliminates the slow and costly process of manual inspections where workers physically patrol roads; secondly, it reduces what used to be days-long reporting processes down to mere seconds; and thirdly, it introduces data-driven decision making where repair crews can objectively pioritize the most dangerous or rapidly worsening defects based on collected evidence rather than guesswork.

   Our smart pothole detection system is designed for continuous improvement through three strategic advancements. The technology will expand its diagnostic scope to identify diverse pavement deteriorations including surface cracks and subsidence areas. Most significantly, advanced pattern recognition algorithms will enable predictive forecasting of potential defect formation, transforming the solution from reactive detection to proactive prevention. These planned enhancements will collectively elevate the systems precision, coverage, and strategic value for infrastructure maintenance. The system is designed to easily plug into cities existing traffic and planning tools, helping fix problems before they become dangerous making roads safer for everyone.

7. REFERENCES

1. Dr. S. N. Gujar, Prathmesh Shinde, Shlok Parihar, Aditya Pawar, Saikrishna Yemul, Pothole Detection System Using YOLO V8,International Research Journal of Modernization in Engineering Technology and Science (IRJMETS), Volume 6, Issue 4, April 2024.

2. Khare, O. M., Gandhi, S., Rahalkar, A. M., Mane, S.,YOLOv8Based Visual Detection of Road Hazards: Potholes, Sewer Covers, and Manholes,arXiv preprint, arXiv:2311.00073, October 2023.

3. Parasnis, S., Mahadik, R., Sarda, M., Narwaria, U.,RoadScan : A Novel and Robust Transfer Learning Framework for Autonomous Pothole Detection in Roads,arXiv preprint, arXiv:2308.03467v2, 2023.

4. Ken Gorro, Elmo Ranolo, Lawrence Roble, Rue Nicole Santillan, Road Pothole Detection Using YOLOv8 with Image Augmentation, Journal of Image and Graphics (JOIG), Volume 12, Number 4, December 2024.

5. Mustafa Yurdakul and S¸akir Tasdemir, An Enhanced YOLOv8 Model for Real-Time and Accurate Pothole Detection and Measurement, arXiv preprint arXiv:2505.04207, May 2025.

6. Uthman Baroudi, Alala BaHamid, Yasser Elalfy, and Ziad Al Alami, Enhancing Pothole Detection and Characterization: Integrated Segmentation and Depth Estimation in Road Anomaly Systems, arXiv preprint arXiv:2504.13648, April 2025.

7. Jiarui Chang, Zhan Chen, and Xia En, RAW-YOLOv8: An Improved YOLOv8 Method for Multi-scale Pothole Detection, In Lecture Notes in Computer Science (LNCS), Springer, Proceedings of Intelligent and Cloud Computing (ICICC 2024), July 2024.

8. Ashur Addanki, Riddhima Srivastava, S. K. Gupta, and R. R. Aditya, Pothole Detection with YOLOv8, ResearchGate, 2023.

9. Naresh Shende, Pawankumar Gupta, Shivam Gupta, Karan Joshi, and Madhav Pawar, Road Pothole Detection Using Deep Learning, International Journal of Scientific Research and Engineering Development (IJSRED), Volume 8, Issue 2, pp. 539544, MarchApril 2025.

10. Dinesh Swami and Mahesh Jangid, Pothole Detection and Prediction Using Deep Learning with CNN and YOLOv8, In Lecture Notes in Networks and Systems (LNNS), Springer, Proceedings of SoCTA 2023, pp. 367379, July 2024.