Visual defect online detection method for traffic facility inspection

The invention discloses a visual defect on-line detection method for traffic facility inspection. The visual defect on-line detection method comprises an off-line training part and an on-line inspection part. During offline training, a defect data set with a bounding box and category labels is adopted as training data, a CrackDet detection model is trained, and the advantages of a deep learning algorithm and a traditional algorithm are combined; meanwhile, different types of defect image blocks are cut out, and a metric learning model is used for training. During on-line inspection, detecting an input video frame by using a detection model to obtain a defect type and a defect position; using a Kalman filter to calculate motion features of the detection frame; using a metric learning model to calculate appearance features of the defect; tracking and counting the detected defects; voting to determine the defect type according to the type detected in the tracking trajectory; finally, online inspection is achieve