Research On Road Abnormal Detection Methods Based On Deep Learning

Abnormalities of the road means that various abnormal diseases appear on the pavement compared to the normal road.Road abnormal detection is an essential and important task in the maintenance of transportation infrastructure.Road damage not only reduces the service life of roads,but also poses serious traffic safety hazards.With the rapid increase in traffic flow,the road network is becoming increasingly diseased.Efficient and accurate road abnormal damage detection can help road maintenance departments to detect and repair in a timely manner to ensure the safety and accessibility.Therefore,we study road damage detection methods with high detection accuracy and real-time performance,and conducts research on road abnormal detection methods based on deep learning methods to provide an important basis and decision support for relevant government departments to maintain road safety.The main research work and results of this paper are as follows:(1)To achieve the object detection task with supervised training requires rich and diverse datasets.We process on the RDDBJ dataset of Beijing urban roads by performing Cut Paste data enhancement,i.e.,cropping and transforming the disease patches and pasting them randomly into the normal images.It not only alleviates the problem of sample imbalance in the dataset,but also helps the model learn the irregularities of abnormal disease targets.The experimental results show that the Cut Paste data enhancement strategy improves the detection accuracy and robustness of the model.(2)To resist the influence of interference information such as shadows and water stains in the background of complex pavement images on the model detection performance,we propose an improved road abnormal detection method based on the YOLOv7 model(referred to as YOLOv7-RDD),introduced the CBAM attention mechanism in the YOLOv7-RDD model,embedded it in the ELAN module of the efficient aggregation network,providing the model extract effective features with attention weights reference,and further improve the stability of the model by using different model ensemble.The experimental results show that the YOLOv7-RDD model can accurately analyze the damage objects under the influence of confounding factors in complex backgrounds and can perform accurate localization.(3)To address the problem of low accuracy of multi-scale damage detection in highresolution pavement images,we propose the model STA-YOLOv7.Based on the attention mechanism of the sliding window design in the Swin-Transformer model,combined with the advantage of hierarchical convolution in the YOLOv7 algorithm to fuse multi-scale features,which can extract global information of targets at different scales,but also maintain fine-grained features.Meanwhile,the dynamic label assignment align Align OTA loss function is introduced training to alleviate the problem of unbalanced sample assignment during sample training and optimize the consistency of single-stage target detector classification and regression.The experimental results demonstrate that the STA-YOLOv7 model can achieve accurate localization in multi-scale targets with high classification accuracy,and can meet the real-time requirements in industrial scenarios,which has application value.