Pothole Detection Using Advanced Neural Networks

Pothole is one of the major types of defects frequently found on the road whose assessment is necessary to process. It is one of the important reason of accidents on the road along with the wear and tear of vehicles. Road defects assessment is to be done through defects data collection and processing of this collected data. Currently, using various types of imaging systems data collection is near about becomes automated but an assessment of defects from collected data is still manual. Manual classification and evaluation of potholes are expensive, labour-intensive, time-consuming and thus slows down the overall road maintenance process. This paper describe a method for classification and detection of the potholes on road images using convolutional neural networks which are deep learning algorithms. In the proposed system we used convolutional neural networks based approach with pre-trained models to classify given input images into a pothole and non-pothole categories. The method was implemented in python using OpenCV library under windows and colab environment, trained on 722 and tested on 116 raw images. The results are evaluated and compared for convolutional neural networks and various seven pre-trained models through accuracy, precision and recall metrics. The results show that pre-trained models InseptionResNetV2 and DenseNet201 can detect potholes on road images with reasonably good accuracy of 89.66%.

Download Full-text

Detecting Potholes in Asphalt Pavement under Small-sample Conditions Based on Improved Faster Region-based Convolution Neural Networks

Canadian Journal of Civil Engineering ◽

10.1139/cjce-2020-0764 ◽

2021 ◽

Author(s):

Lili Pei ◽

Li Shi ◽

Zhaoyun Sun ◽

Wei Li ◽

Yao Gao ◽

...

Keyword(s):

Neural Networks ◽

Asphalt Pavement ◽

Small Sample ◽

Small Samples ◽

Detection Accuracy ◽

Fusion Model ◽

Convolution Neural Networks ◽

Image Set ◽

Pothole Detection ◽

Different Sources

Pavement potholes have low detection accuracy under the condition of small samples. To address this issue, we propose a method for efficient and accurate pothole detection under small-sample conditions, based on improved Faster R-CNN (Region-based Convolution Neural Networks). First, images consisting of different pothole shapes and sizes are acquired from different sources and then, augmented and denoised to obtain the image set. Second, two representative target detection models, Faster R-CNN and YOLOv3, are tested. The detection results indicate that Faster R-CNN achieves better detection performance. Furthermore, to overcome inconsistencies (missed detections and inaccurate position estimations), the feature extraction layers of VGG16, ZFNet, and ResNet50 networks are used in combination with Faster R-CNN. The results show that the VGG16+Faster R-CNN fusion model yields superior accuracy. Finally, the detection accuracy improved to 0.8997 after adjusting the size of the candidate frame, which also enabled the successful detection of previously missed targets.

Download Full-text