Pavement Pothole Detection Based on Cascade and Fusion Convolutional Neural Network Using 2D Images under Complex Pavement Conditions

Abstract

Background: Background: The development of deep learning technology has promoted the industrial intelligence, and automatic driving vehicles have become a hot research direction. As to the problem that pavement potholes threaten the safety of automatic driving vehicles, the pothole detection under complex environment conditions is studied. Objective: The goal of the work is to propose a new model of pavement pothole detection based on convolutional neural network. The main contribution is that the Multi-level Feature Fusion Block and the Detector Cascading Block are designed and a series of detectors are cascaded together to improve the detection accuracy of the proposed model. Methods: A pothole detection model is designed based on the original object detection model. In the study, the Transfer Connection Block in the Object Detection Module is removed and the Multi-level Feature Fusion Block is redesigned. At the same time, a Detector Cascading Block with multi-step detection is designed. Detectors are connected directly to the feature map and cascaded. In addition, the structure skips the transformation step. Results: The proposed method can be used to detect potholes efficiently. The real-time and accuracy of the model are improved after adjusting the network parameters and redesigning the model structure. The maximum detection accuracy of the proposed model is 75.24%. Conclusion: The Multi-level Feature Fusion Block designed enhances the fusion of high and low layer feature information and is conducive to extracting a large amount of target information. The Detector Cascade Block is a detector with cascade structure, which can realize more accurate prediction of the object. In a word, the model designed has greatly improved the detection accuracy and speed, which lays a solid foundation for pavement pothole detection under complex environmental conditions.

Road pothole detection based on improved YOLOv7

“…From the results, it can be seen that compared with YOLOv3, SSD, and Faster R-CNN models, it has increased by 16.02%, 24.02%, and 41.52%. Compared with the improved YOLOv5 model used in reference [12] , it increases by 7.78%.…”

“…The model used mAP% Reference [10] YOLOv3 67% Reference [10] SSD 59% Reference [11] Faster R-CNN 41.5% Reference [12] Improved YOLOv5 75.24% The paper Improved YOLOv7 83.02%…”