Automated identification of containers based on visual features

^{Tom Fisk / Pexels}

The ContainerVision project aims to optimize the management of containers at smaller inland and seaports. To this end, aerial images are to be analyzed in order to identify the containers visible on them so that port employees are supported in their search for a specific container.

Until now, containers had to be searched for manually. This is very time-consuming and the use of container stackers and other vehicles results in high fuel consumption. Our project automatically evaluates aerial images from drones, for example, using computer vision methods. In many cases, it will not be possible to read the identification numbers of the containers correctly due to the quality of the images and the distance, so that manual identification is required. For this reason, we capture and process a variety of visual features such as color, logos, external defects and surface structure. This significantly simplifies drone flights, enabling more frequent scans and thus a higher temporal resolution.

The implementation of the project comprises several steps. The starting point is data collection, followed by data preparation in order to construct a database of container characteristics. Using an image processing pipeline developed by us, the containers can then be identified, with the aforementioned visual features serving as classification parameters.

Subsequently, the focus is on the recognition of specific containers. The captured information is compared with a database in order to narrow down the candidates for possible correct identification numbers and ultimately to uniquely identify a container. The database is supplemented with the characteristics of newly recorded containers. This automated identification of containers using visual features does not yet exist.

In concrete terms, the image processing pipeline could look as follows: After loading the image, scaling, noise reduction and contrast adjustment are performed to improve the image quality. In addition, a color calibration process is carried out in conjunction with histogram equalization to make images comparable under different lighting conditions. This is followed by container detection and segmentation with the aim of localizing the containers in the image and separating them from their background.

The containers are then classified according to their visual characteristics, which are obtained using conventional analytical and machine learning methods. This includes, among other things, recording the identification number using optical character recognition (OCR). The algorithms are trained in advance with a sufficient amount of data in the form of containers with corresponding labels. The condition of the container, for example with regard to possible defects and colors, can be determined by image segmentation, image comparison, object tracking and pattern recognition. After classification, the individual containers are identified based on the classification results using decision trees or neural networks. In this way, the ContainerVision project contributes to more efficient port logistics and to an improvement in economic and ecological sustainability. In addition, ContainerVision promotes the use of drone technology, image processing and machine learning in the logistics industry.

And here the group in their own words (German):