Comparative Analysis of YOLO-Based Object Detection Models for Road Damage Detection Us- ing UAV Imagery

Comparative Analysis of YOLO-Based Object Detection Models for Road Damage Detection Us- ing UAV Imagery

Ms. Dola Jyothi, G. Chandrika, M. Jahnavi, M. Sakshitha

CSE(AI&ML) CMR Institution Of Technology, Hyderabad, Telangana, India

Abstract – Road damage detection plays a critical role in ensuring transportation safety and effective infrastructure maintenance; however, conventional manual inspection methods are labor-intensive, time-consuming, and suscep- tible to human error. The use of aerial imagery provides scalable monitoring capabilities, yet accurately identifying cracks and potholes under varying illumination, texture complexity, and scale variations remains a significant chal- lenge. This paper presents a comparative analysis of three real-time object detection models for automated road damage detection using aerial images. The proposed sys- tem incorporates dataset preprocessing, structured data splitting, model training, and quantitative performance evaluation using accuracy, precision, recall, and F1-score metrics. Experimental results demonstrate that the most advanced detection model achieves an accuracy of 92%, precision of 94%, recall of 90%, and an F1-score of 92%, outperforming the other evaluated models while maintain- ing efficient inference performance. The findings confirm that advanced deep learning-based object detection frameworks provide a reliable and scalable solution for automated road infrastructure monitoring, enabling im- proved maintenance planning and enhanced transporta- tion safety.

Keywords: Road damage detection, aerial image analysis, deep learning, object detection, real-time monitoring, in- frastructure inspection

1. INTRODUCTION

   manual surveys or vehicle-mounted equipment, which are labor-intensive, time-consuming, and often inconsistent due to subjective assessment.

   Recent advancements in aerial imaging and deep learning have introduced automated solutions for infrastructure moni- toring. Unmanned aerial platforms enable large-scale image acquisition with high spatial resolution, while convolutional neural network-based object detection models allow automat- ed localization and classification of defects. Among these, single-stage real-time detection frameworks have gained prominence due to their balance between computational effi- ciency and detection accuracy. Despite these advancements, several challenges remain. Existing studies often evaluate a single detection model without systematic comparison across architectures. Furthermore, limited work investigates perfor- mance trade-offs between detection accuracy, recall, and real- time feasibility in aerial road inspection scenarios.

   To address these gaps, this paper presents a structured comparative analysis of three real-time object detection archi- tectures for automated road damage detection using aerial imagery. The proposed framework integrates dataset prepro- cessing, model training, quantitative evaluation, and perfor- mance benchmarking under consistent experimental condi- tions.

   The main contributions of this work are summarized as fol- lows:

   Road transportation networks form the backbone of eco- nomic development and public mobility. However, road sur- face deterioration such as cracks, potholes, and surface wear significantly affects vehicle safety, travel comfort, and maintenance costs. According to global infrastructure reports, delayed detection of road damage increases repair expenses by up to 40% due to progressive structural degradation. Tradi- tional road inspection methods rely on

   • A unified experimental framework for evaluating multiple real-time object detection architectures on aerial road damage datasets.

   • A quantitative performance comparison using accu- racy, precision, recall, F1-score, and mean average precision metrics.

   • An analysis of detection robustness under varying environmental and scale conditions.

   • Identification of the most efficient architecture bal- ancing detection performance and inference speed for practical deployment.

   The results demonstrate that optimized deep learning detection frameworks can significantly improve automated infrastruc- ture monitoring and support data-driven maintenance deci- sion-making.

2. RELATED WORK

   Automated road damage detection has evolved significantly with advances in computer vision and deep learning. Existing research can be broadly categorized into three themes: tradi- tional image-processing approaches, two-stage deep learning detectors, and real-time single-stage detection frameworks.

   1. Traditional Image Processing Approaches

      Early studies relied on handcrafted feature extraction tech- niques such as edge detection, thresholding, morphological operations, and texture analysis for crack identification. These approaches were computationally lightweight and interpreta- ble. However, they were highly sensitive to illumination changes, shadows, and noise. While such methods performed adequately in controlled environments, their generalization to complex outdoor scenarios was limited, especially for irregu- lar pothole structures and varying pavement textures.

   2. Two-Stage Deep Learning Detectors

      The introduction of convolutional neural networks signifi- cantly improved defect detection accuracy. Two-stage detec- tors, such as region-based convolutional frameworks, first generate region proposals and then classify them. These mod- els achieve high localization precision and perform well on small object detection. However, their multi-stage pipeline increases computational overhead, making real-time aerial deployment challenging. Although these methods handle complex feature extraction effectively, they struggle to meet the speed requirements of large-scale infrastructure monitor- ing.

   3. Single-Stage Real-Time Detection Frameworks

      Single-stage object detection models address the speed lim- itations of two-stage approaches by performing localization and classification in a unified architecture. These models offer a favorable balance between detection accuracy and inference speed, making them suitable for real-time applications. Recent architectural improvements focus on enhanced feature pyra- mids, better backbone efficiency, and optimized loss functions to improve small-object detection and robustness under scale variations.

      While recent studies report strong detection performance using single-stage architectures, several limitations remain. First, many works evaluate only one detection model without

      systematic comparison across architectures under identical experimental settings. Second, performance metrics are often reported inconsistently, making objective benchmarking diffi- cult. Third, limited studies analyze the trade-off between de- tection accuracy and real-time inference capability in aerial road inspection scenarios.

   4. Comparative Analysis and Research Gap

   Table I summarizes representative approaches in automated road damage detection and highlights their relative strengths and limitations.

   Approach Type	Detection Accuracy	Inference Speed	Compu- tational Cost	Limitations
   Tradi- tional Im- age Pro- cessing	Moder- ate	High	Lw	Sensitive to lighting and noise
   Two- Stage Deep Learning	High	Low	High	Not suitable for real-time UAV deploy- ment
   Single- Stage De- tection	High	High	Mod- erate	Performance varies architec- tures

   Table 1:comparative analysis

   While prior work demonstrates strong performance in iso- lated scenarios, a structured comparative analysis under uni- fied experimental conditions remains limited. In particular, systematic evaluation of multiple real-time detection architec- tures using consistent datasets and metrics is underexplored.

   To address this gap, the present study conducts a controlled comparative analysis of three real-time object detection archi- tectures for aerial road damage detection. By benchmarking accuracy, precision, recall, F1-score, and mean average preci- sion under identical training conditions, this work provides clearer insights into architectural trade-offs and practical de- ployment feasibility.

3. METHODOLOGY

   This section describes the dataset characteristics, prepro- cessing pipeline, model configuration, training setup, and evaluation protocol used in this study. All experimental pro- cedures were conducted under controlled and consistent con- ditions to ensure reproducibility.

   1. Dataset Description

      The dataset consisted of high-resolution aerial images cap- turing road surfaces with visible defects, including longitudi- nal cracks, transverse cracks, potholes, and surface wear. A total of 3,200 images were collected using an unmanned aerial platform equipped with a 20-megapixel RGB camera. Images were captured at an average altitude of 812 meters under daylight conditions.

      Ground-truth annotations were generated using bounding box labeling in YOLO format. Each defect instance was as- signed a class label corresponding to its damage type. The dataset was divided into training (70%), validation (15%), and testing (15%) subsets using stratified sampling to preserve class distribution

   2. Data Preprocessing

      All images were resized to 640 × 640 pixels to ensure uni- form model input dimensions. Pixel values were normalized to the range [0,1]. Data augmentation techniques were applied during training to improve generalization, including:

      • Horizontal flipping (probability 0.5)

      • Random brightness adjustment (±15%)

      • Small-angle rotation (±10 degrees)

      • Random scaling (0.81.2 factor)

        No augmentation was applied during validation or testing. Bounding box coordinates were automatically adjusted af-

        ter geometric transformations.

   3. Model Architectures

      Fig no 1: System Architecture

      Architecture illustrates the overall architecture of the pro- posed road damage detection framework. The system consist- ed of six major components:

      1. Data Acquisition

         Raw UAV images were captured using an aerial imaging platform and stored in a centralized database. The UAV col- lected high-resolution RGB images of road surfaces contain- ing cracks and potholes. The acquired images were transmit- ted to the storage module for further processing.

      2. Data Preprocessing

         The raw UAV images underwent preprocessing before model training. This module performed:

         • Image normalization

         • Resizing to fixed input dimensions (640 × 640)

         • Data cleaning

         • Annotation formatting

           The output of this stage was a set of normalized images suitable for deep learning model input.

      3. YOLO Model Layer

         The processed images were passed to three detection archi- tectures:

         • YOLOv5

         • YOLOv7

         • YOLOv8

           Each model independently performed:

         • Feature extraction using a convolutional backbone

         • Multi-scale feature aggregation

         • Bounding box regression

         • Damage classification

           The outputs included bounding box coordinates, confidence scores, and damage class predictions.

      4. Evaluation Module

         The predictions generated by each YOLO model were for- warded to the evaluation module. This module computed quantitative performance metrics, including:

         • Accuracy

         • Precision

         • Recall

         • F1-score

           These metrics were calculated based on comparisons be- tween predicted bounding boxes and ground-truth annotations using Intersection-over-Union thresholding.

      5. Database and Storage

         All intermediate and final results were stored in a central- ized database. This included:

         • Raw UAV images

         • Normalized images

         • Trained model weights

         • Detection outputs

         • Performance metrics

           The storage module enabled systematic experiment tracking and reproducibility.

      6. Storage and Interface Layer

         The final module provided visualization and performance reporting through a graphical interface. Users could:

         • View detected road damage with bounding boxes

         • Compare YOLO model performance

         • Analyze metric summaries

      This layer ensured practical usability for infrastructure monitoring applications.

   4. Training Configuration

      Training was performed on a workstation equipped with an NVIDIA RTX 3060 GPU (12 GB memory), 16 GB RAM, and

      an Intel i7 processor.

      The following hyperparameters were used:

      • Batch size: 16

      • Number of epochs: 50

      • Initial learning rate: 0.001

      • Optimizer: Adam

      • Weight decay: 0.0005

      • Confidence threshold: 0.25

      • Intersection-over-Union threshold: 0.5

        The learning rate was reduced by a factor of 0.1 if valida- tion loss plateaued for five consecutive epochs.

   5. Evaluation Metrics

      Model performance was evaluated using the following quantitative metrics:

      Mean Average Precision at Intersection-over-Union thresh- old 0.5 (mAP@0.5) was used as the primary detection metric.

      Statistical comparison between models was performed us- ing paired t-tests with significance level set at p < 0.05.

   6. Algorithm and Logic

   Input: UAV images of road surfaces Output: Detected road damages with location and classifica- tion

   Step 1: Image Acquisition

   Capture aerial images of roads using UAV and store them in the dataset.

   Step 2: Data Preprocessing

   Resize all images to a fixed size and normalize pixel values. Apply data augmentation techniques such as flipping, rotation, and brightness adjustment to improve model robustness.

   Step 3: Dataset Splitting

   Divide the dataset into training, validation, and testing sets to ensure proper model evaluation.

   Step 4: Model Initialization

   Intialize the YOLO models with predefined parameters such as batch size, number of epochs, and learning rate.

   Step 5: Model Training

   Train the YOLO models using the training dataset. During training, the model learns to identify patterns and features associated with different types of road damage.

   Step 6: Detection Process

   Pass test images into the trained model. The model predicts bounding boxes, class labels, and confidence scores for de- tected damages.

   Step 7: Result Filtering

   Remove detections with low confidence scores and elimi- nate duplicate detections using Non-Maximum Suppression to improve accuracy.

   Step 8: Performance Evaluation

   Compare predicted results with ground truth annotations and calculate evaluation metrics such as accuracy, precision, recall, F1-score, and mAP.

   Step 9: Model Comparison

   Compare the performance of YOLOv5, YOLOv7, and YOLOv8 models and identify the best-performing model based on accuracy and inference time.

   Step 10: Final Output

   Display the detected road damages with bounding boxes and labels, and store the results for further analysis.

4. RESULT ANALYSIS

   1. Detection Output Visualization

      Figure 2 shows a sample detection output generated using the YOLOv8 model. The model localized the damaged road region with a bounding box and classified it as Road Dam- aged. The detection confidence exceeded the predefined threshold of 0.25.

      Fig1: YOLOv8 detection output

      Fig3 presents the detection output obtained using the YOLOv7 model. Multiple damage regions were identified in a single image. The model successfully detected both crack pat- terns and pothole structures under varying illumination condi- tions.

      Fig 2: YOLOv8 detection output illustrating multiple detected damage regions.

      Fig4 illustrates the detection output generated using the YOLOv8 model. The model detected complex pothole struc- tures and irregular crack boundaries with improved localiza- tion alignment relative to ground-truth annotations.

      Fig3: YOLOv8 detection output showing bounding box predictions on aerial road imagery.

   2. Quantitative Performance Comparison

      Table I summarizes the performance metrics obtained on the test dataset.

      Table I Performance Metrics Comparison

      Model	Accuracy (%)	Precision (%)	Recall (%)	F1-Score (%)
      YOL Ov5	88.2 ± 1.5	89.4 ± 1.3	86.7 ± 1.6	88.0 ± 1.4
      YOL Ov7	91.6 ± 1.2	92.3 ± 1.1	90.8 ± 1.3	91.5 ± 1.2
      YOL Ov8	95.2 ± 1.0	94.8 ± 0.9	93.9 ± 1.1	94.3 ± 1.0

      Table 2 :performance metrics

      YOLOv8 achieved the highest mean accuracy of 95.2% (±1.0%). The difference between YOLOv8 and YOLOv5 was

      7.0 percentage points. Statistical testing using paired t-tests indicated that the improvement was significant (p < 0.01).

   3. Performance Graphical Representation

      Figure 8 presents a bar chart comparing accuracy, precision, recall, and F1-score across the three models.

      The bar chart clearly shows a progressive increase in per- formance metrics from YOLOv5 to YOLOv8. Error bars rep- resent standard deviation across five independent runs.

      Fig8. Comparative bar chart of performance metrics for YOLOv5, YOLOv7, and YOLOv8.

   4. Inference Time Analysis

   Average inference time per image was measured across 500 test samples.

   Model	Inference Time (ms)
   YOLO v5	18 ms
   YOLO v7	21 ms
   YOLO v8	16 ms

   Table 3: inference

   YOLOv8 recorded the lowest average inference time while maintaining the highest detection performance.

   .

5. DISCUSSION

   This study aimed to determine whether a structured com- parative evaluation of real-time object detection architectures could identify a model that improves road damage detection accuracy while maintaining computational efficiency. The experimental results demonstrated that YOLOv8 achieved the highest detection accuracy (95.2%), precision (94.8%), recall (93.9%), and F1-score (94.3%), while also recording the low- est inference time (16 ms per image). These findings directly address the research objective by showing that improved ar- chitectural design can enhance detection performance without compromising real-time capability.

   The observed performance improvement may be attributed to enhanced feature aggregation mechanisms and refined bounding box regression strategies implemented in the most recent architecture. The higher recall value suggests improved sensitivity to small and irregular damage regions, while the increased precision indicates reduced false positive detections. Although YOLOv7 demonstrated competitive performance, its inference latency was higher than YOLOv8, suggesting a trade-off between architectural complexity and computational efficiency.

   Compared to traditional image-processing techniques re- ported in earlier studies, which were sensitive to lighting and texture variations, the deep learning-based models evaluated in this work demonstrated improved robustness under com- plex aerial imaging conditions. Similarly, while two-stage detection frameworks have previously achieved high localiza- tion accuracy, their computational overhead limits practical deployment for real-time UAV monitoring. The results of this study align with recent research trends emphasizing single- stage architectures for infrastructure inspection tasks.

   Despite these promising findings, several limitations must be acknowledged. First, the dataset size, although balanced, was limited to controlled UAV capture conditions. Perfor- mance may vary under extreme weather, nighttime imaging, or different pavement materials. Second, the evaluation was

   conducted using a fixed Intersection-over-Union threshold of 0.5; alternative thresholds could influence metric values. Third, hardware constraints may affect inference time measurements across different computational platforms.

6. CONCLUSION

   This study presented a structured comparative analysis of three real-time object detection models for UAV-based road damage detection. The results demonstrated that the most re- cent architecture achieved the highest detection performance, reaching 95.2% accuracy while maintaining the lowest infer- ence time of 16 ms per image. These findings directly answer the research question by confirming that architectural im- provements can enhance detection precision and recall without compromising real-time processing capability.

   The proposed evaluation framework provided a consistent benchmarking environment, enabling objective comparison across models using standardized metrics. The results indicate that advanced single-stage detection architectures offer a prac- tical and scalable solution for automated road infrastructure monitoring. By combining high detection accuracy with com- putational efficiency, the proposed approach supports real- time deployment in intelligent transportation and maintenance systems.

7. FUTURE WORK

   1. Integration with Advanced YOLO Versions Future work can explore newer versions such as YOLOv9 or upcoming architectures to improve detection accuracy, speed, and robustness, especially in complex environments.

   2. Real-Time Deployment on Edge Devices The model can be optimized and deployed on edge devices such as smartphones, drones, or embedded systems (e.g., Raspberry Pi). This will enable real-time road damage detec- tion without relying on high-end computing resources.

   3. Expansion of Dataset The performance of the model can be improved by training it on a larger and more diverse dataset that includes different road conditions, weather scenarios, lighting variations, and geographical regions.

   4. Multi-Class Damage Classification Future systems can classify different types of road damage such as cracks, potholes, surface wear, and patches more accu- rately using fine-grained labeling techniques.

   5. Integration with GIS and Smart City Systems The detection system can be integrated with Geographic In- formation Systems (GIS) to map detected damages and assist municipal authorities in efficient maintenance planning.

   6. Automated Alert and Maintenance System An automated alert system can be developed to notify authori-

      ties instantly when road damage is detected, enabling faster response and repair.

   7. Improving Model Efficiency Further research can focus on reducing model size and compu- tational cost using techniques such as model pruning, quantiza- tion, and knowledge distillation.

   8. Video-Based Detection Enhancement Instead of relying only on images, future work can include video stream processing for continuous monitoring and more accurate detection using temporal information.

   9. User-Friendly Mobile Application A mobile app can be developed for public use, allowing users to capture and upload images of road damage, contributing to a crowd-sourced monitoring system.

   10. Integration with Autonomous Vehicles The system can be extended to assist autonomous vehicles in detecting road conditions and adapting driving behavior ac- cordingly.

8. REFERENCES

1. A. Bochkovskiy, C. Wang, and H. Liao, YOLOv4: Optimal Speed and Accuracy of Object Detection, arXiv preprint arXiv:2004.10934, 2020.

   Available: https://arxiv.org/abs/2004.10934

2. G. Jocher et al., YOLOv5, 2020. Available: https://github.com/ultralytics/yolov5

3. Z. Zou, Z. Shi, Y. Guo, and J. Ye, Object Detection in 20 Years: A Survey, IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 43, no. 8, Aug. 2021.

4. C. Wang, A. Bochkovskiy, and H. Liao, YOLOv7: Trainable Bag-of- Freebies Sets New State-of-the-Art for Real-Time Object Detectors, arXiv preprint arXiv:2207.02696, 2022. Available: https://arxiv.org/abs/2207.02696

5. G. Jocher et al., YOLOv8, 2023. Available: https://github.com/ultralytics/ultralytics

6. A. Ultralytics, Ultralytics YOLO Documentation and Releases, 2024. Available: https://docs.ultralytics.com