Enhanced Road Manhole Cover Detection Using MGB-YOLO: A Deep Learning Approach

By Dr. Sampath LonkaReviewed by Susha Cheriyedath, M.Sc.Oct 3 2023

In a recent publication in the journal Scientific Reports, researchers introduced a real-time method for detecting manhole covers on roads based on deep learning models.

Background

Road manhole covers play a pivotal role in urban infrastructure, granting access to vital underground utility networks such as gas, sewage, and water systems. If neglected or poorly marked, they pose substantial traffic hazards. In the context of traffic safety, recognizing the vehicle environment swiftly enhances safety. While deep learning thrives in identifying road participants such as vehicles, pedestrians, and signs, the recognition of manhole covers lags. Addressing this gap is vital for traffic safety.

Deep learning models for object detection

With the advancement of artificial intelligence and deep learning, object detection techniques have made significant progress, especially in traffic environments. Numerous researchers have conducted studies focused on detecting objects on roads to ensure safer traffic operations. The application of deep learning methods in analyzing traffic videos contributes to the overall improvement of roadway safety. Several works in this field include vehicle detection and tracking, even in adverse weather conditions; road anomaly detection in autonomous vehicles; a real-time system for traffic signs and road object detection; and detecting road manhole covers.

The You Only Look Once (YOLO) series of object identification algorithms has grown in popularity because of their precision and quick performance. YOLO model variants excelled at several tasks, including dual-lane detection, 3D object detection, and sign recognition. They cannot, however, successfully identify small, closely packed items.

The architecture of the MGB-YOLO model

The model YOLOv5 demonstrates superior performance in terms of speed and computational resource utilization across various datasets. Researchers chose YOLOv5s, the smallest version of the YOLOv5 series, to strike a compromise between performance, speed, and model size. Its architecture includes three key components: the backbone network, the neck network, and the detection network. The backbone extracts image features using convolutional layers, while the neck network integrates these features from different scales, enhancing semantic information and positional data. The detection network predicts object categories and locations.

The enhancement of the backbone with MobileNet-V3, a lightweight neural network architecture, improves performance on resource-constrained devices. This version incorporates hard-swish activation and adaptability for small models. Replacing the final layer (C3 module) in YOLOv5s with the Global Attention Mechanism (GAM) enhances the model's ability to process contextual information and improves object detection, especially for small or densely placed objects.

The BottleneckCSP architecture is introduced to the neck network, offering a balance between information flow and computational efficiency. It reduces parameters, maintains precision, and speeds up inference, benefiting object detection. The loss function in YOLOv5s considers bounding box regression, confidence, and target classification, with a focus on mitigating false positives for small targets such as road manhole covers.

Researchers proposed a deep learning model called MGB-YOLO, combining YOLOv5s with MobileNet-V3, GAM, and BottleneckCSP. This model caters to resource-constrained devices while achieving efficient and accurate manhole cover detection on road surfaces. MobileNet-V3 enhances efficiency, GAM improves contextual understanding, and BottleneckCSP streamlines information flow. The result is a lightweight and high-performance network model for this specific application.

Performance and comparison analysis

Image Data Acquisition and Augmentation: A unique dataset for road manhole covers was curated, consisting of 1000 images showcasing standard and damaged covers. The dataset fosters a comprehensive understanding of both normal and damaged conditions. To combat data scarcity issues, five augmentation techniques were employed: original images, Gaussian blur, brightness adjustment, noise addition, and 90-degree rotation. This expanded the dataset from 1000 to 6000, enhancing model robustness.

Experimental Results: The experiment evaluates the model's training process and object detection effectiveness using parameters such as box loss, objectness loss, class loss, precision, recall, and mean average precision (mAP) from training and testing datasets. Loss values decrease rapidly in the first 100 epochs, stabilizing after 200 epochs. The model, trained on qualified and unqualified manhole covers, achieves a satisfactory overall testing performance.

Comparisons with Different Models: The proposed MGB-YOLO model outperforms YOLO v5s, Single Multi-Frame Detector (SSD), Faster Region-Convolutional Neural Network (Faster-RCNN), YOLOv7, and YOLOv8s in terms of mAP, model size, and detection speed. MGB-YOLO achieves a smaller model size and a higher mAP, surpassing other models in detection precision. Although its mAP improvement over YOLOv7 and YOLOv8s is modest, its smaller storage footprint makes it suitable for in-vehicle embedded devices.

Classification Performance: Confusion matrices show MGB-YOLO's superior performance in correctly classifying both qualified and unqualified manhole covers, even outperforming Faster-RCNN and SSD. For background detection, all models exhibit relatively low accuracy due to diverse background shapes, a common challenge in small target detection.

Practical Detection Results: MGB-YOLO accurately identifies five manhole covers, demonstrating universality and applicability. It outperforms other algorithms in detecting the same objects, especially small targets. In detecting unqualified manhole covers, MGB-YOLO reduces miss detection and error rates compared to SSD and Faster-RCNN. It provides a higher accuracy level for small targets and effectively reduces omission and error rates, indicating its superiority.

Conclusion

In summary, researchers proposed an enhanced model, MGB-YOLO, tailored for road manhole cover detection. By incorporating MobileNet-V3, GAM, and BottleneckCSP, the model strikes a balance between detection precision and computational speed. Comparative experiments with other methods reveal MGB-YOLO's superior recognition precision and model efficiency. Future research will concentrate on enhancing real-time detection, putting the model in use on embedded devices for vehicles, and fine-tuning data augmentation for improved accuracy.