Research On Recognition And Location Technology Of Container Keyhole Based On Machine Vision

With the rapid development of international trade,container transportation,as the mainstream mode of transportation of goods,has grown rapidly.Automated terminals need container transportation to unmanned and intelligent development,and container loading and unloading operation is a key part of container transportation.The time spent on the location of the spreader and container keyhole will directly affect the production efficiency of the terminal.At present,most domestic terminals still rely on the experience of operators to achieve the alignment of the spreader and the container keyhole.It not only has high work intensity,but also easily causes safety accidents and affects the economic benefits of the terminal.Therefore,this thesis focuses on the recognition and location of container keyhole based on machine vision,including the research and comparison of the recognition method of container keyhole,and the research on the precise location technology of the keyhole area obtained after recognition.The main research contents of this paper are as follows:1.Production of container keyhole samples.Since there is no public data set on the network of the container keyhole,this paper collects the image of the container at the dock site,and then combines the OpenCV function library to manually mark and normalize the keyhole part in the container image,which lays a foundation for the model training later.2.Design and implementation of container keyhole recognition algorithm based on HOG+SVM and LBP+AdaBoost.Aiming at the problem of container keyhole recognition,this paper proposes a HOG-based and SVM classifier and a recognition algorithm based on LBP feature and AdaBoost classifier.A data set consisting of a positive sample image containing a keyhole and a negative sample image containing no keyhole is used as a training set for subsequent classifiers.After completing the feature extraction and classifier training,the multi-scale sliding window is used to identify the keyhole.Aiming at the problem that the detection time in the initial HOG+SVM keyhole recognition is too long,the recognition algorithm is improved,and the global scanning of each frame image is avoided,which reduces the detection time.Finally,the two recognition algorithms are analyzed and compared.3.Design and implementation of container keyhole recognition algorithm based on convolutional neural network.In order to improve the real-time and accuracy of container keyhole recognition,this paper designs a container keyhole recognition algorithm based on convolutional neural network and improves the convolutional neural network: According to the actual keyhole detection requirements,the anchor points of different scales are used to ensure the detection effect while taking into account the real-time detection.The labeled keyhole samples were iteratively trained through three different layer network models.After comparison and analysis,the deep convolutional neural network detection model based on YOLOv3 was obtained,and the threshold parameters were adjusted to test the effect of container keyhole recognition.4.The realization of precise location technology for container keyhole.According to the initial recognition of the keyhole area,the necessary image enhancement processing is performed on the area,and then the contour of the keyhole is extracted.The fitting of the keyhole of the container is completed by the obtained keyhole contour information,thereby further improving the accuracy of the container keyhole location.5.Experimental analysis of container keyhole recognition and location algorithm.In this paper,the improved HOG+SVM container keyhole recognition algorithm,LBP+AdaBoost based container keyhole algorithm and YOLOv3-based container keyhole recognition algorithm are tested.The recognition algorithm based on YOLOv3 is more ideal in real-time and accuracy.The trained model is in the test set.The accuracy rate is 97.6%,the recall rate is 83.1%,F1 is 89.8%,and the detection time of a single image on the GPU is 11.21 ms.In the experiment of locating the center of container keyhole,for the image sequence of the video capture pixel of 1292*964,the error of the center of the keyhole during the daytime is within 15 pixels,and the error of the center of the lock hole at night is within 20 pixels.The container keyhole recognition algorithm based on YOLOv3 combined with the subsequent positioning algorithm can accurately locate the container keyhole.