Technology About Container Stereo Recognition And Location Based On Deep Learning

With the development of container operation towards automation,intellectualization and unmanned,the demand for container hoisting operations of ports is increasing.The intelligent operation system based on machine vision is the future direction of development of port equipment,which uses the visual inspection technology to deal with the problem of container recognition and location.Due to the changeable weather and harsh coastal environment of port,the background of actual container target has a lot of irrelevant information,and the traditional target detection and recognition algorithm is limited by the illumination,weather and scenes of the port,it has created challenges and difficulties in port container detection and recognition.In recent years,with the development of visual detection and recognition technology,this topic aims to improve level of the container automated stereo recognition and location,and has carried out the theoretical analysis and experimental research to its two key technologies-container detection and recognition technology based on deep learning and container stereo positioning technology based on binocular vision.The main work is as follows:(1)Aiming at to meet the research requirements for container stereo recognition and location,an experimental platform of container stereo recognition and location system has been built.The hardware has completed the design and selection of the vision system.The algorithm has researched idea of container recognition based on Faster R-CNN and container depth location based on binocular vision.(2)The images of datasets with complex environment backgrounds which were collected by manual acquisition,amplification and labeling,then container datasets with strong robustness were constructed for training and testing of models.This paper needed to use deep learning model for container target recognition,it was necessary to train the model effectively to achieve the best effect,so the accuracy and optimization of container datasets was particularly important.In addition,a semi-automatic annotation method based on GrabCut segmentation algorithm was proposed,which can segment foreground and background of target to produce large-scale container datasets quickly.(3)Aiming at the problem of scale change,complex background,light change and weather disturbance during container operation,an improved Faster R-CNN method was proposed.By reducing the parameters of the model network,adding adversarial spatial transformer network,enhancing sample target foreground features,multi-scale training and learning strategies,the training efficiency and recognition performance of the network were improved.The improved model was tested and compared,and experiments showed that the method can effectively recognition the container target.(4)For the high-precision stereo positioning problem of containers,a measurement and positioning method based on binocular vision was proposed according to the analysis of 3D vision measuring technology.Stereo matching was performed after the calibration and correction,and a pyramid stereo matching network was proposed to get the matching cost which makes the matching error rate of the final disparity map result lower.Finally,we calculated the depth information and got the three-dimensional coordinate of the container,then carried out experiments to verify the accuracy of container recognition and location algorithm under different conditions.