Surface Vessel: Local Positioning System Based on Computer Vision

Object detection is a technology capable of finding instances of objects in images. Among the main applications in many different areas the following can be highlighted: traffic monitoring, robots and security. A typical object detection algorithm takes an image as an input, finds the object of interest in it, and provides a bounding box. The output bounding box is usually axis-aligned, and it is enough for many applications. However, there are situations when the angle of rotation of the bounding box is essential and should be calculated. In this paper, different approaches for local positioning of a single object with obtainment of center coordinates and direction angle are studied and tested. Firstly, a basic image processing approach that uses binarization by the color range of the object of interest is considered. Furthermore, this simple algorithm finds the final bounding box with an angle of rotation using geometrical properties of the surface vessel. This approach obtains angle by getting the minimum bounding box using the contour of the surface vessel and principal component analysis. In addition, deep learning segmentation approaches were considered. Considering all algorithms, the image processing approach based on threshold binarization performs faster than deep learning approaches, suiting into the problem of local positioning of a surface vessel in a suitable way. Moreover, it provides comparable accuracy and execution time with the deep learning segmentation approaches. The final algorithm obtained can detect the surface vessel and obtain its linear (chi, gamma) and angular (a) coordinates. It is robust to the appearance of other objects with the same colors as the object of interest and of smaller size. It is also robust to the appearance of objects of any size not in the color range of the object of interest.