Exécution de notre outil de synthèse sur un nouvel article

Résumé : Road obstacles are one of the most significant factors leading to traffic accidents, particularly in the transportation scenes such as tunnels. Due to the current target detection technology requiring datasets with category labels for training, there are significant limitations, this paper overcomes this problem and proposes a novel method for detecting abnormal objects such as debris or scattered objects on road surface in tunnel. The proposed method can achieve non-category detection of abnormal objects on the tunnel road surface, which is detecting the existence of objects without classifying them. This is implemented by using image semantic segmentation as abnormal objects such as debris or scattered objects always occupy a few pixels in an image, and the essence of semantic segmentation is to classify targets at the pixel level, so semantic segmentation is very suitable for detecting small targets. We explained in detail the principle of our proposed method, mainly including the schematic of our proposed method, the dataset preparation, model selecting and training. The preliminary experimental results on our test dataset show that the average precision and average recall of detection are both greater than 60%, which has a great potential for real practical applications and has great promotion value.